<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[My blog's RSS Feed]]></title><description><![CDATA[Always Awake]]></description><link>https://blogik.netlify.app</link><generator>GatsbyJS</generator><lastBuildDate>Mon, 29 Mar 2021 00:45:04 GMT</lastBuildDate><item><title><![CDATA[욕심많은 알파카(김성익)]]></title><description><![CDATA[상세 한양대학교 정보시스템학과 재학(2017~) 한양대학교 개발동아리 Forif 활동(2017)) 웹 프로그래밍 연합동아리 피로그래밍 12기 활동(2020) 웹 프로그래밍 연합동아리 피로그래밍 13기 회장(2020) AWSKRUG 대학생 그룹 AUSG…]]></description><link>https://blogik.netlify.app/about/</link><guid isPermaLink="false">https://blogik.netlify.app/about/</guid><content:encoded>&lt;h1 id=&quot;상세&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%83%81%EC%84%B8&quot; aria-label=&quot;상세 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;상세&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;한양대학교 정보시스템학과 재학(2017~)&lt;/strong&gt;&lt;br/&gt;
&lt;strong&gt;한양대학교 개발동아리 Forif 활동(2017))&lt;/strong&gt;&lt;br/&gt;
&lt;strong&gt;웹 프로그래밍 연합동아리 피로그래밍 12기 활동(2020)&lt;/strong&gt;&lt;br/&gt;
&lt;strong&gt;웹 프로그래밍 연합동아리 피로그래밍 13기 회장(2020)&lt;/strong&gt;&lt;br/&gt;
&lt;strong&gt;AWSKRUG 대학생 그룹 AUSG 4기 활동(2020.09~2021.08)&lt;/strong&gt;&lt;br/&gt;
&lt;strong&gt;네이버 부스트캠프 AI Tech 1기(2021.01~2021.06)&lt;/strong&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[모델 경량화 4 - 행렬 분해(Filter Decomposition)]]></title><description><![CDATA[행렬 분해 by 홍원의 마스터님, BoostCamp AI Tech 8주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/48_filter_decompostion/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/48_filter_decompostion/</guid><pubDate>Fri, 19 Mar 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;행렬-분해low-rank-approximation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%96%89%EB%A0%AC-%EB%B6%84%ED%95%B4low-rank-approximation&quot; aria-label=&quot;행렬 분해low rank approximation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;행렬 분해(Low-rank approximation)&lt;/h1&gt;&lt;p&gt;행렬 분해라고 했지만, &lt;strong&gt;&lt;code&gt;Low-rank approximation&lt;/code&gt;&lt;/strong&gt;은 단순히 행렬(matrix)만 다루는 것이 아니기 때문에 &lt;strong&gt;저차원(저랭크) 분해(decomposition)&lt;/strong&gt; 정도로 볼 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Filter / Kernel / Matrix decompostion&lt;/li&gt;&lt;li&gt;Matrix / Tensor / Low-rank Factorization&lt;/li&gt;&lt;li&gt;Low-rank Approximation&lt;/li&gt;&lt;li&gt;위 용어들은 대부분 비슷한 의미로 사용된다.&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;three-maps&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#three-maps&quot; aria-label=&quot;three maps permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Three maps&lt;/h2&gt;&lt;h3 id=&quot;map-1--matrixtensor-is-a-data-modeling-tool&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#map-1--matrixtensor-is-a-data-modeling-tool&quot; aria-label=&quot;map 1  matrixtensor is a data modeling tool permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Map 1 : Matrix(Tensor) is a data modeling tool&lt;/h3&gt;&lt;p&gt;전체 구성은 이렇다.&lt;/p&gt;&lt;p&gt;&lt;img src=&quot;./img/lightweight/Map1.png&quot; alt=&quot;Map1&quot;/&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A_{nn}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 행렬은 &lt;strong&gt;Eigenvalue(Spectral) decompostion&lt;/strong&gt;과 &lt;strong&gt;Diagonalization&lt;/strong&gt;으로 나타낼 수 있다.&lt;/li&gt;&lt;li&gt;정사각 행렬이 아닌 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A_{mn}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;행렬로 일반화하면, &lt;strong&gt;Singular value decompostion&lt;/strong&gt;이 된다.&lt;/li&gt;&lt;li&gt;이 개념을 행렬이 아니라 3차원 이상의 Tensor까지 확장하면 &lt;strong&gt;CP decompostion&lt;/strong&gt;이나 &lt;strong&gt;Tucker decompostion&lt;/strong&gt;을 적용시킬 수 있다(Tucker decompostion이 좀 더 일반적인 방식이다).&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;map-2--matrixtensor-is-a-linear-transformationmap&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#map-2--matrixtensor-is-a-linear-transformationmap&quot; aria-label=&quot;map 2  matrixtensor is a linear transformationmap permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Map 2 : Matrix(Tensor) is a linear transformation(map)&lt;/h3&gt;&lt;p&gt;선형변환(map)이란, 차원 축의 이동을 의미한다.&lt;/p&gt;&lt;p&gt;좀 더 formal한 definition으로는, 변환 전과 변환 후의 덧셈과 scalar를 보존하는 변환을 의미한다.&lt;/p&gt;&lt;h3 id=&quot;map-3--terminology&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#map-3--terminology&quot; aria-label=&quot;map 3  terminology permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Map 3 : terminology&lt;/h3&gt;&lt;p&gt;강의 슬라이드를 참조하자. 선형대수학적인 용어들이 연결되어있다. 특히 노란색 박스는 꼭 숙지하자.&lt;/p&gt;&lt;h3 id=&quot;gaussian-elimination&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gaussian-elimination&quot; aria-label=&quot;gaussian elimination permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Gaussian elimination&lt;/h3&gt;&lt;p&gt;가우스 소거법은 행렬의 1차방정식(row)들을 적절하게 조작하여 미지수를 소거하고, 남은 미지수에 대한 선형 결합(1차 방정식)들로 표현하는 기법이다. 이 과정에서 남은 1차 방정식들을 &lt;strong&gt;&lt;code&gt;basis(기저)&lt;/code&gt;&lt;/strong&gt;라고 할 수 있다. 기저는 공간 축을 나타낸다.&lt;/p&gt;&lt;p&gt;n차원 실수 공간 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;R&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathbb{R}^n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68889em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.664392em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 하위 공간인 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;S&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 있다고 하자. &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;S&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 벡터들의 집합인데, 이 벡터들은 1. &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;S&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 span이면서 2.선형독립적이다. 1번은 해당 벡터가 기저의 변형이라는 말이고, 2번은 그 벡터가 내포한 원형 기저가 다른 벡터의 원형 기저와 다른 기저라는 의미이다. 겹치면(dependant) 동일한 기저가 될테니, 사실상 의미없는 벡터가 될 것이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;잘 이해가 되지 않으면, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x+2y=2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.66666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8388800000000001em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;4&lt;/mn&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;4&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;2x+4y=4&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8388800000000001em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;4&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 생각해보자. 두 식은 분명 다른 식이지만, 두 식의 기저, 즉 공간은 동일하다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;가우스 소거법을 행하면서 벡터를 소거하기 때문에, &lt;strong&gt;기존의 rank를 축소하게 된다.&lt;/strong&gt;&lt;/p&gt;&lt;h2 id=&quot;kernel-method&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#kernel-method&quot; aria-label=&quot;kernel method permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Kernel method&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Kernel&lt;/code&gt;&lt;/strong&gt;이란 말은 원래 알맹이, 핵심을 의미한다. 대수학에서는 V→W로 이동할 때 W의 중심으로 매핑되는 원소들의 집합을 커널이라고 한다. OS에도, 이미지 프로세싱에도 동일한 용어가 사용되어서 혼선이 있다(umbrella term). 느슨하게 말하자면 central essential part라고 볼 수 있다.&lt;/p&gt;&lt;p&gt;&lt;img src=&quot;https://miro.medium.com/max/700/1*mCwnu5kXot6buL7jeIafqQ.png&quot; alt=&quot;https://miro.medium.com/max/700/1*mCwnu5kXot6buL7jeIafqQ.png&quot;/&gt;&lt;/p&gt;&lt;p&gt;기존의 차원에서는 classify하기 힘든 정보들이, 차원을 높이면 조금 더 분리하기 쉬워질 수 있다. 위의 이미지에서처럼 2차원의 데이터들을 3차원 공간으로 이동시키면, 이들을 가르는 더 명확한 decision surface를 얻을 수 있다(항상 그런것은 아니다).&lt;/p&gt;&lt;p&gt;그러나, 이처럼 차원을 높여 분리하면, 두 점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x,y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 간의 거리를 얻기 위해 dot product를 할 때 늘어난 차원의 벡터도 추가로 계산해야 하여 연산량이 과중해진다(computationally expensive).&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Kernel method&lt;/code&gt;&lt;/strong&gt;는 이러한 문제점을 막기 위한 방법으로, 고차원에서 dot product한 결과를 가지고 있는 함수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;K(x,y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 토대로 결과를 내되, 기존의 차원에서 dot product하여 연산을 줄이는 기법이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Essential한 part는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;K(x,y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 되고, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;K(x,y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서 이미 구한 부분은 불필요하니 구하지 않겠다는 방식이다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;low-rank-approximation-in-model-compression&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#low-rank-approximation-in-model-compression&quot; aria-label=&quot;low rank approximation in model compression permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Low-rank approximation in model compression&lt;/h3&gt;&lt;p&gt;ML 모델에서 filter를 decompose한다.&lt;/p&gt;&lt;div&gt;&lt;p&gt;kernel이 아니라 filter임을 유의하자. kernel은 (WxH)의 2D array이고, filter는 채널까지 포함한 3D array를 일컫는다.&lt;/p&gt;&lt;/div&gt;&lt;p&gt;Depth-wise separable Convolution을 통하여 원래의 Regular Convolution에 비하여 Computing 비용을 줄여준다. 다만 Kernel method처럼 원래의 결과와 완전히 동일하지는 않고, approximate하다.&lt;/p&gt;&lt;h3 id=&quot;low-rank-tensor-approximation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#low-rank-tensor-approximation&quot; aria-label=&quot;low rank tensor approximation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Low-rank tensor approximation&lt;/h3&gt;&lt;p&gt;기존의 모델은 텐서로 이루어져있다. 텐서 모델을 low-rank로 decompose하여 낮은 랭크의 텐서들로 분해시켜 근사(approximate)한다. 이 때 분해하여 얻어낸 텐서들은 essential part이므로 일종의 kernel이라고 볼수도 있다. Convolution의 그 커널과는 다른 의미이다.&lt;/p&gt;&lt;p&gt;대표적인 data-free 방법이다. 이미 학습된 모델을 추가데이터 없이 분해하여 사용하는 방식이기 때문이다.&lt;/p&gt;&lt;h2 id=&quot;matrix-decompostion&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#matrix-decompostion&quot; aria-label=&quot;matrix decompostion permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Matrix decompostion&lt;/h2&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n\times m&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.66666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 matrix를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(n\times r)\cdot(r\times m)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;으로 분해한다. 또는 matrix를 잘라내어서(truncate) 사용하기도 한다.&lt;/p&gt;&lt;h3 id=&quot;eigenvalue-decompostion&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#eigenvalue-decompostion&quot; aria-label=&quot;eigenvalue decompostion permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Eigenvalue Decompostion&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;고유벡터(Eigenvalue)&lt;/code&gt;&lt;/strong&gt;란, 어떤 선형변환을 취했을 때 방향은 바뀌지 않고 크기만 변하는 벡터를 의미한다. &lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:29.6875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAAAwElEQVQY003PSZKEMAxEUd//kgQb5hnMWND9otiUF46U/DMlh7+fc57nMAxd16VpOs9zjDHLsiRJlmVZ13XbNh1Cue87Pvya7/v+fM91XUo3A05oWZZC3XVd53luxvM8oe97ecKgNNTMaZoYjuOA8ojQbNsWoGkGgA5vi1kXN46jVG9KBqPM8apZFIUsa+/bXlVVXGNg8IAzmTCZUyKIQPs28ZqbpqExhH2DH97f45O0n7gRVrW5RKgg6FsCXpLlH8JiV+q+05I6AAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;eigenvalue&quot; title=&quot;eigenvalue&quot; src=&quot;/static/c2f3a94c50f6d3c1d0207633303f97b8/2bef9/eigenvalue.png&quot; srcSet=&quot;/static/c2f3a94c50f6d3c1d0207633303f97b8/6f3f2/eigenvalue.png 256w,/static/c2f3a94c50f6d3c1d0207633303f97b8/01e7c/eigenvalue.png 512w,/static/c2f3a94c50f6d3c1d0207633303f97b8/2bef9/eigenvalue.png 1024w,/static/c2f3a94c50f6d3c1d0207633303f97b8/71c1d/eigenvalue.png 1536w,/static/c2f3a94c50f6d3c1d0207633303f97b8/fde66/eigenvalue.png 1574w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;실수이면서 대칭인 (정방)행렬 A가 있다면, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mrow&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P^{-1}=P^T&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8141079999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8413309999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 성립한다.&lt;/p&gt;&lt;h3 id=&quot;singular-value-decompostionsvd&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#singular-value-decompostionsvd&quot; aria-label=&quot;singular value decompostionsvd permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Singular Value Decompostion(SVD)&lt;/h3&gt;&lt;p&gt;Eigenvalue Decompostion의 케이스를 그대로 들고와서, 정방행렬이 아닌 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n\times m&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.66666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;행렬에도 적용시킬 수 있다. &lt;/p&gt;&lt;p&gt;슬라이드를 다시 읽어보자.&lt;/p&gt;&lt;h3 id=&quot;truncated-svd-⇒-principal-component-analysispca&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#truncated-svd-%E2%87%92-principal-component-analysispca&quot; aria-label=&quot;truncated svd ⇒ principal component analysispca permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;(Truncated) SVD ⇒ Principal Component Analysis(PCA)&lt;/h3&gt;&lt;p&gt;&lt;code&gt;PCA&lt;/code&gt;는 SVD에서 한걸음 더 나아가서, 기존의 SVD에서 비교적 더 중요한 eigenvalue들, 즉 principal component들만 남기고 다 제거해버리는것이다.&lt;/p&gt;&lt;p&gt;이외에도 CPD, NMF, TKD 등의 여러 방법들이 있다.&lt;/p&gt;&lt;h2 id=&quot;tensor-decompostion&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#tensor-decompostion&quot; aria-label=&quot;tensor decompostion permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Tensor decompostion&lt;/h2&gt;&lt;p&gt;기존의 모델은 텐서로 이루어져있다. 텐서 모델을 low-rank로 decompose하여 낮은 랭크의 텐서들로 분해시켜 근사(approximate)한다. 이 때 분해하여 얻어낸 텐서들은 essential part이므로 일종의 kernel이라고 볼수도 있다. Convolution의 그 커널과는 다른 의미이다.&lt;/p&gt;&lt;p&gt;대표적인 data-free 방법이다. 이미 학습된 모델을 추가데이터 없이 분해하여 사용하는 방식이기 때문이다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[모델 경량화 2 - Quantization(양자화)]]></title><description><![CDATA[양자화 by 홍원의 마스터님, BoostCamp AI Tech 8주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/46_quantization/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/46_quantization/</guid><pubDate>Thu, 18 Mar 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;quantization&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#quantization&quot; aria-label=&quot;quantization permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Quantization&lt;/h1&gt;&lt;h2 id=&quot;fixed-point--floating-point&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#fixed-point--floating-point&quot; aria-label=&quot;fixed point  floating point permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Fixed-point &amp;amp; Floating-point&lt;/h2&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:22.65625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAFCAIAAADKYVtkAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA9ElEQVQY0zWOW2+CMABG+f8/Zw9L9kIYQ1bEjktBRUYdYEsLBXRObmqyumlycvLlPH0KjvbIE3nainLPioaxhpTfpDzS8vhvSVH9sLqTLsRJxviL46zK2UFBpo1gDN589QU4ILJBokH+6lQS3RWaHK54NvCTulYXzPBb3aulZ6jVnFqhZOD8QulE/kbXXYbpwXjzOF374dz147086MezElorCBLLWM/NzeI99pxs5hamx02fmX5p+txwmWZnOtzpkIBAgKCywnq+bEBQK1m83aUVcuVhz4OrLeaYnDDp7tB+kx4+wtxZErRhmN4i+hRBItyo/AWOcAdrbIjoTgAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;fixed-floating&quot; title=&quot;fixed-floating&quot; src=&quot;/static/66c441f2960e67956b9fc9326043d728/2bef9/fixed-floating.png&quot; srcSet=&quot;/static/66c441f2960e67956b9fc9326043d728/6f3f2/fixed-floating.png 256w,/static/66c441f2960e67956b9fc9326043d728/01e7c/fixed-floating.png 512w,/static/66c441f2960e67956b9fc9326043d728/2bef9/fixed-floating.png 1024w,/static/66c441f2960e67956b9fc9326043d728/71c1d/fixed-floating.png 1536w,/static/66c441f2960e67956b9fc9326043d728/92bb4/fixed-floating.png 1824w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;Fixed Point&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;정수부와 실수부를 나누어 표기한다.&lt;/li&gt;&lt;li&gt;소수 표기 등이 쉽고, floating point에 비하여 훨씬 더 빠르고 효율적이다.&lt;/li&gt;&lt;li&gt;그러나 커버하는 범위가 적다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Floating Point&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;지수부와 가수부로 나눈다.&lt;/li&gt;&lt;li&gt;자릿수를 바꿀 때, 지수부만 변경하면 된다. 따라서 정확성을 얼마 잃지 않고 훨씬 더 넓은 숫자 범위를 커버할 수 있다. &lt;strong&gt;&lt;div&gt;단, 정보의 표현능력은 같다(2^bit개만 표현할 수 있다)&lt;/div&gt;&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;대부분의 시스템에서 사용된다.&lt;/li&gt;&lt;li&gt;FPU(Floating Point Unit)은 정수나 fixed-point 연산보다 하드웨어(특히 ARM-Mobile/IoT)에서 구현하기 훨씬 더 어렵다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;회로 공간이나 메모리 요구량에서 Integer보다 floating-point number가 더 많은 리소스를 요구한다.&lt;/p&gt;&lt;h3 id=&quot;precision--accuracy&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#precision--accuracy&quot; aria-label=&quot;precision  accuracy permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Precision &amp;amp; Accuracy&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;Precision : 분산(variance)개념. 얼마나 몰려있느냐?&lt;/li&gt;&lt;li&gt;Accuracy : 편향(bias)개념. 목표치에서 얼마나 벗어나 있느냐?&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;사격으로 치면 영점이 잘맞는 것이 Accuracy가 높은 것, 탄착군 형성 시 Precision이 높다고 생각할 수 있다.&lt;/p&gt;&lt;h2 id=&quot;quantization이란&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#quantization%EC%9D%B4%EB%9E%80&quot; aria-label=&quot;quantization이란 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Quantization이란&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;모델 사이즈 축소(reduction) - 표현능력은 조금 줄어든다.&lt;/li&gt;&lt;li&gt;메모리 bandwidth 요구치를 맞추는 데에 도움이 된다.(감소시켜서)&lt;/li&gt;&lt;li&gt;on-device의 int8 연산이 float32(FLU 사용) 연산보다 더 빠르다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;양자화(Quantization)&lt;/code&gt;&lt;/strong&gt;란, 추론 속도를 높이는 테크닉을 주로 말한다(양자 연산은 forward pass만 지원한다).&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;div&gt;즉, 주 목적은 training time을 줄이는게 아니라 inference time을 줄이는 것이다.&lt;/div&gt;&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:47.65625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;quantization-lossy-conversion&quot; title=&quot;quantization-lossy-conversion&quot; src=&quot;/static/8b7aca25c611ed1c9194d04889ca8552/2bef9/quantization-lossy-conversion.png&quot; srcSet=&quot;/static/8b7aca25c611ed1c9194d04889ca8552/6f3f2/quantization-lossy-conversion.png 256w,/static/8b7aca25c611ed1c9194d04889ca8552/01e7c/quantization-lossy-conversion.png 512w,/static/8b7aca25c611ed1c9194d04889ca8552/2bef9/quantization-lossy-conversion.png 1024w,/static/8b7aca25c611ed1c9194d04889ca8552/f2f8c/quantization-lossy-conversion.png 1490w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;양자화는 float32를 최대 256개까지밖에 표현할 수 없는 int8형태로 바꾸는 것이기 때문에, lossy conversion이다. 원래의 정보를 어느정도 잃을 수 밖에 없다.&lt;/p&gt;&lt;p&gt;DL은 지금까지 가장 나은 예측(highest precision)을 보여주기 위한 방법으로 사용해왔는데, 그렇다면 굳이 양자화하는 등의 low precision을 사용해야하는 필요성은 무엇인가?&lt;/p&gt;&lt;p&gt;→ 딥러닝은 연산량이 굉장히 많아 알고리즘으로 치면 효율이 아주 좋지 않은 기법이다. 너무 많은 파라미터를 학습해야 하므로, 어느정도 정보를 손실하고서라도 이를 줄일 가치가 있다.&lt;/p&gt;&lt;h3 id=&quot;affine-quantization&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#affine-quantization&quot; aria-label=&quot;affine quantization permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Affine quantization&lt;/h3&gt;&lt;p&gt;우리가 지금까지 배워왔던 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y=\sigma(wx+b)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 형태의 변환을 &lt;code&gt;Affine 변환&lt;/code&gt;이라고 한다. Affine 변환의 특징은, 변환 전 x와 y의 차이가 변환후 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f(x)-f(y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 차이로 그대로 맵핑될 수 있다는 것이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;m_f(x-y) = f(x)-f(y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;이 경우, distance는 그대로 보존되지 않지만, distance의 ratio는 보존된다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;quantized-value-미분&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#quantized-value-%EB%AF%B8%EB%B6%84&quot; aria-label=&quot;quantized value 미분 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;quantized value 미분&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:53.515625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAALCAIAAADwazoUAAAACXBIWXMAABYlAAAWJQFJUiTwAAABWUlEQVQoz21S246CMBDt//+XjxtNTLw8mOiiCJSCUEovQ4sHuqvu6qQ005meOcPpsOPxeD6fkyTB3jSNc87ONsCIom+MwWl8M2ZejIi89yF4JLIs41zAQQR7mqZfs61Wq+VyuV6vhRDsvR4pG4bQqV42eV2hRDoGdTp9LxaL7Xa72Wz2uz1WyUs2+NDbiTA2pm9W1WYmG70RojiU+cGpPIwf7IcZDTsiXVvT2BgJYbqPilobKRWarOtaax1TMcsAg0cDtVwZ6R6w3wo+TCp8JB6ZlNLRUF1v4pQpnrdZ1hUFUJfLpSyFc/Rgm+WnV9kZavaVNa3GuwSojUaIkJCyU7Jum6qqyjBoYNB2nuc4P8FKaNc56EUz5tEwPttdeborrnsrk0jOOceLPsFOTf+JSYjt/TME+76PioIcI+RfJGCYKtV1CP1lnmYDMBSFgx3XIHjbtkopkMcSd4sEfOjCdBXUAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;differentiating-quantized-values&quot; title=&quot;differentiating-quantized-values&quot; src=&quot;/static/5b5557eee60cc9e7b121ce025f8535a3/2bef9/differentiating-quantized-values.png&quot; srcSet=&quot;/static/5b5557eee60cc9e7b121ce025f8535a3/6f3f2/differentiating-quantized-values.png 256w,/static/5b5557eee60cc9e7b121ce025f8535a3/01e7c/differentiating-quantized-values.png 512w,/static/5b5557eee60cc9e7b121ce025f8535a3/2bef9/differentiating-quantized-values.png 1024w,/static/5b5557eee60cc9e7b121ce025f8535a3/71c1d/differentiating-quantized-values.png 1536w,/static/5b5557eee60cc9e7b121ce025f8535a3/07d37/differentiating-quantized-values.png 1790w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;그런데 문제는 forward pass하면서 quantization하다보면, backward pass시에 미분불가능한 점들이 생겨버리므로 역전파가 불가능해진다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{\partial y}{\partial x}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.277216em;vertical-align:-0.345em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9322159999999999em&quot;&gt;&lt;span style=&quot;top:-2.6550000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.446108em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.345em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 quantize 되기 전의 값으로 집어넣어 계산한다. 그러면 그냥 항등함수라 미분이 제대로 되지는 않지만, 일단 backward pass를 통과할수는 있다.&lt;/li&gt;&lt;li&gt;또는, quantization이 안된 값과 완전 quantization 된 값들의 중간을 취해 부드러운 계단식으로 만들어서(smoothing) 미분가능하게 하여 통과시키는 방법도 있다.&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;quantization의-종류&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#quantization%EC%9D%98-%EC%A2%85%EB%A5%98&quot; aria-label=&quot;quantization의 종류 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Quantization의 종류&lt;/h2&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:44.921875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;quantization-category&quot; title=&quot;quantization-category&quot; src=&quot;/static/7c38b3d3e7f0e41e708cba516a289937/2bef9/quantization-category.png&quot; srcSet=&quot;/static/7c38b3d3e7f0e41e708cba516a289937/6f3f2/quantization-category.png 256w,/static/7c38b3d3e7f0e41e708cba516a289937/01e7c/quantization-category.png 512w,/static/7c38b3d3e7f0e41e708cba516a289937/2bef9/quantization-category.png 1024w,/static/7c38b3d3e7f0e41e708cba516a289937/71c1d/quantization-category.png 1536w,/static/7c38b3d3e7f0e41e708cba516a289937/569c6/quantization-category.png 1674w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;How to quantize?&lt;ul&gt;&lt;li&gt;Dynamic : weight만 quantization하고 있다가 inference 시점에만 activation quantize함&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;When to quantize?&lt;ul&gt;&lt;li&gt;Post-training(PTQ, static) : 학습 이후에 quantization 하는것&lt;/li&gt;&lt;li&gt;quantization-aware training(QAT) : 학습 과정에서 quantization을 가정하고 시뮬레이션을 같이 돌림&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;하드웨어(CPU, GPU, TPU 등)나 라이브러리마다 지원하는 quantization 종류가 다르다. 그 때 맞추어 적절하게 사용해야한다.&lt;/p&gt;&lt;h2 id=&quot;quantization-결과-테이블-읽기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#quantization-%EA%B2%B0%EA%B3%BC-%ED%85%8C%EC%9D%B4%EB%B8%94-%EC%9D%BD%EA%B8%B0&quot; aria-label=&quot;quantization 결과 테이블 읽기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Quantization 결과 테이블 읽기&lt;/h2&gt;&lt;hr/&gt;&lt;p&gt;강의 자료가 굉장히 유용하다. 더 정확하게 이해하고싶다면 PDF의 reference를 모두 들어가 보자.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[모델 경량화 3 - 지식 증류(Knowledge Distillation)]]></title><description><![CDATA[지식 증류 by 홍원의 마스터님, BoostCamp AI Tech 8주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/47_knowledge_distillation/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/47_knowledge_distillation/</guid><pubDate>Thu, 18 Mar 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;knowledge-distillation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#knowledge-distillation&quot; aria-label=&quot;knowledge distillation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Knowledge Distillation&lt;/h1&gt;&lt;h2 id=&quot;knowledge&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#knowledge&quot; aria-label=&quot;knowledge permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Knowledge&lt;/h2&gt;&lt;p&gt;&lt;img src=&quot;http://www.allthingy.com/wp-content/uploads/2014/07/Wisdom-Knowledge-Information-Data-Pyramid15.png&quot; alt=&quot;http://www.allthingy.com/wp-content/uploads/2014/07/Wisdom-Knowledge-Information-Data-Pyramid15.png&quot;/&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Data → Information → Knowledge → Wisdom순으로 Data가 정제된다.&lt;/li&gt;&lt;li&gt;Knowledge distillation은 이 중 knowledge를 뽑아내는 기술이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;추가 레퍼런스&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://opentutorials.org/module/3653/22995&quot;&gt;[logit과 sigmoid, softmax의 관계]&lt;/a&gt;&lt;/p&gt;&lt;h2 id=&quot;knowledge-distillation-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#knowledge-distillation-1&quot; aria-label=&quot;knowledge distillation 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Knowledge distillation&lt;/h2&gt;&lt;p&gt;소금물에서 증류하여 소금을 얻어내듯이, &lt;strong&gt;&lt;code&gt;지식 증류(knowledge distillation)&lt;/code&gt;&lt;/strong&gt;는 커다란 Teacher 모델에서 엑기스(지식)만 뽑아내어 작은 Student 모델로 전달하는 방식이다.&lt;/p&gt;&lt;p&gt;Transfer Learning과 Knowledge Distillation의 차이는 이렇게 비유할 수 있다. 지식을 전달한다는 점에서는 둘 모두 동일하지만, 지식 증류는 사이즈를 줄이는데에 목적이 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Transfer learning - ex) 영어에서 배운 지식을 프랑스어에 적용시키는 것(도메인이 다름)&lt;/li&gt;&lt;li&gt;Knowledge distillation - ex) 선생님이 학생에게 역사 지식을 가르쳐주는 것(도메인은 같음)&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;teacher-student-networks--hinton-loss&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#teacher-student-networks--hinton-loss&quot; aria-label=&quot;teacher student networks  hinton loss permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Teacher-Student networks &amp;amp; Hinton loss&lt;/h2&gt;&lt;p&gt;전체적인 과정은 다음과 같다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:50%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;knowledge-distillation-diagram&quot; title=&quot;knowledge-distillation-diagram&quot; src=&quot;/static/4991b7f69b0a541cfe9524e047e76da7/2bef9/knowledge-distillation-diagram.png&quot; srcSet=&quot;/static/4991b7f69b0a541cfe9524e047e76da7/6f3f2/knowledge-distillation-diagram.png 256w,/static/4991b7f69b0a541cfe9524e047e76da7/01e7c/knowledge-distillation-diagram.png 512w,/static/4991b7f69b0a541cfe9524e047e76da7/2bef9/knowledge-distillation-diagram.png 1024w,/static/4991b7f69b0a541cfe9524e047e76da7/71c1d/knowledge-distillation-diagram.png 1536w,/static/4991b7f69b0a541cfe9524e047e76da7/5df5d/knowledge-distillation-diagram.png 1572w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; = 1일 경우 익히 알고있는 softmax function이므로, 가장 큰 값(즉 가장 확신하는 값)을 제외하고는 다 버린다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 1보다 큰 어떤 상수가 될 경우 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 soften 되어 &amp;#x27;덜 확신하는&amp;#x27; 값이 된다.(soft prediction).&lt;ul&gt;&lt;li&gt;정답이라고 생각하는 값만 중요한 게 아니라, 확신의 과정에서 어느 클래스와 어느 클래스가 더 비슷하다고 생각했는지같은 것도 모두 정보의 일종이다. 이러한 개념을 제프리 힌튼 교수가 처음으로 제시했다.(&lt;em&gt;&amp;quot;The relative probablities of the incorrect outputs tell us a lot about how the model tned to generalize.&amp;quot;&lt;/em&gt;)&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;자세한 설명은 &lt;a href=&quot;https://blogik.netlify.app/BoostCamp/U_stage/38_annotation_data_efficient_learning/#knowledge-distillation&quot;&gt;[이전의 글]&lt;/a&gt;을 참조해보자.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 상수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 어떻게 바뀌느냐에 따라 이렇게 바뀌게된다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:50%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAKCAIAAAA7N+mxAAAACXBIWXMAABYlAAAWJQFJUiTwAAABHElEQVQoz4VSAW7DIAzM/584VVubpCRpO1LABpsEMpN00bpO2slCBvt8tkXFzNM0zfMsZ1ox/wXiCD5ssTQnQADEiohyzssKYkZECS8vcGDExNmSJe1uTLUJShlr3alu3g6Hs1J7uRXFB6c9GFHdrjllIVais6XGGAE8ohg+k5fSNnnvHXPcH0WycgDLf5BR7zC6YEMI+1CFLAV/5v3S3MBMBsY7jojuSfm77VfO+phTWU+waDToq4Vx30Ihy6i7Wl7xWkfflJPlatu37yUjpQcZnDvXH5dzrZpj39VD36j22DWnq2o+u1YP3VXVelCyZ6YwDqo9HfStf5CJyQcM5IlCMSYxuXoC9NYa7dHEKa7fJ8v/kJA0K8PGOH0BBkhH+48e1J8AAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;softmax-t&quot; title=&quot;softmax-t&quot; src=&quot;/static/7dcd968acef84abc5d3dbb67e01226fe/2bef9/softmax-t.png&quot; srcSet=&quot;/static/7dcd968acef84abc5d3dbb67e01226fe/6f3f2/softmax-t.png 256w,/static/7dcd968acef84abc5d3dbb67e01226fe/01e7c/softmax-t.png 512w,/static/7dcd968acef84abc5d3dbb67e01226fe/2bef9/softmax-t.png 1024w,/static/7dcd968acef84abc5d3dbb67e01226fe/71c1d/softmax-t.png 1536w,/static/7dcd968acef84abc5d3dbb67e01226fe/a878e/softmax-t.png 2048w,/static/7dcd968acef84abc5d3dbb67e01226fe/33714/softmax-t.png 3336w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h2 id=&quot;zero-mean-assumption&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#zero-mean-assumption&quot; aria-label=&quot;zero mean assumption permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Zero-mean assumption&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;Zero-mean assumption&lt;/code&gt;&lt;ul&gt;&lt;li&gt;logit값들이 zero-mean이 아닌 distillation은 Model compression이라고 부르기 어렵다는 것을 의미한다.&lt;/li&gt;&lt;li&gt;유도 과정은 &lt;a href=&quot;https://drive.google.com/file/d/1N8lOUoAAIodT7IuRIRRG5i9I2Ltpbxh9/view?usp=sharing&quot;&gt;[여기]&lt;/a&gt;에 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;즉, distillation ≠ compression임을 이야기한다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[모델 경량화 1 - Pruning(가지치기)]]></title><description><![CDATA[가지치기 by 홍원의 마스터님, BoostCamp AI Tech 8주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/45_pruning/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/45_pruning/</guid><pubDate>Wed, 17 Mar 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;pruning-for-network-compression&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#pruning-for-network-compression&quot; aria-label=&quot;pruning for network compression permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Pruning for network compression&lt;/h1&gt;&lt;p&gt;모델 학습 시 중요한 파라미터는 살리고 그렇지 않은 파라미터는 덜어내는(가지치기하는) 경량화 테크닉이다.&lt;/p&gt;&lt;h2 id=&quot;weighted-sum&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#weighted-sum&quot; aria-label=&quot;weighted sum permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Weighted sum&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;가중 평균.&lt;/strong&gt;&lt;/p&gt;&lt;p&gt;&lt;em&gt;Decision theory&lt;/em&gt;에서, 중요한 것은 더 반영하고, 중요하지 않은 것은 덜 반영하는 것을 의미한다. 즉, 가중치를 두어서 평균을 내는것(value에 반영하는것)을 의미한다.&lt;/p&gt;&lt;h2 id=&quot;pruning이란&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#pruning%EC%9D%B4%EB%9E%80&quot; aria-label=&quot;pruning이란 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Pruning이란?&lt;/h2&gt;&lt;p&gt;&lt;img src=&quot;https://miro.medium.com/max/1400/1*rw2zAHw9Xlm7nSq1PCKbzQ.png&quot; alt=&quot;https://miro.medium.com/max/1400/1*rw2zAHw9Xlm7nSq1PCKbzQ.png&quot;/&gt;&lt;/p&gt;&lt;p&gt;Neural Network뿐만 아니라 decision tree에서도 많이 사용되는 방법이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;장점&lt;ul&gt;&lt;li&gt;추론 속도가 빨라진다(파라미터가 줄어드므로)&lt;/li&gt;&lt;li&gt;Regularization(모델 복잡도를 줄인다)이 일반화 성능을 높인다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;단점&lt;ul&gt;&lt;li&gt;정보의 손실이 생긴다&lt;/li&gt;&lt;li&gt;입자도(granularity, 세밀함)가 하드웨어 가속 디자인의 효율성에 영향을 미친다.&lt;ul&gt;&lt;li&gt;너무 sparse하게 만들어버리면 하드웨어 가속 효율이 떨어진다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:28.90625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;pruning_graph&quot; title=&quot;pruning_graph&quot; src=&quot;/static/cf20586d924cee451f63cf63220c19ae/2bef9/pruning_graph.png&quot; srcSet=&quot;/static/cf20586d924cee451f63cf63220c19ae/6f3f2/pruning_graph.png 256w,/static/cf20586d924cee451f63cf63220c19ae/01e7c/pruning_graph.png 512w,/static/cf20586d924cee451f63cf63220c19ae/2bef9/pruning_graph.png 1024w,/static/cf20586d924cee451f63cf63220c19ae/71c1d/pruning_graph.png 1536w,/static/cf20586d924cee451f63cf63220c19ae/a878e/pruning_graph.png 2048w,/static/cf20586d924cee451f63cf63220c19ae/62b11/pruning_graph.png 3438w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위의 사진을 보면 pruning 이후에 0주위의 weight distribution이 대폭 축소(outline이 10^5→10^4)된 것을 알수 있다. 이는 별로 의미가 없다고 여겨지는 0주위의 정보들을 다 잘라내는 것이다.&lt;/p&gt;&lt;p&gt;그런데 pruning과 이전에 배웠던 dropout 기법은 일견 비슷해 보인다. pruning과 dropout은 무엇이 다른가?&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Dropout도 regularization을 위해 쓴다.&lt;/li&gt;&lt;li&gt;Dropout은 앙상블 효과를 낼 수 있다. 즉, 어떤 뉴런을 하나 끄고 학습할 때마다 서로 다른 structure의 네트워크를 여러 개 중복해서 학습하는 효과가 있다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Pruning&lt;/code&gt;은 한번 잘라낸 뉴런을 보관하지 않는다&lt;/strong&gt;. 그러나 &lt;strong&gt;&lt;code&gt;Dropout&lt;/code&gt;은 regularization이 목적이므로 학습 시에 뉴런들을 랜덤으로 껐다가 (보관해두고) 다시 켜는 과정을 반복한다.&lt;/strong&gt; 추론 시에는 모든 뉴런을 켜고 수행한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;과정은 [pruning으로 mask 업데이트 → fine-tune으로 weight 업데이트]를 반복적으로 수행함으로써 이루어진다.&lt;/p&gt;&lt;h3 id=&quot;regularization&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#regularization&quot; aria-label=&quot;regularization permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Regularization&lt;/h3&gt;&lt;p&gt;&lt;img src=&quot;./img/lightweight/regularizaiton.png&quot; alt=&quot;regularizaiton&quot;/&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;randomly initialized weight &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서 시작해서 training을 한다.&lt;/li&gt;&lt;li&gt;만약 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\lambda = 0&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;λ&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이라면, regularization을 수행하지 않는다는 말이다. 그러면 파란색 점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta^*&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.688696em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mbin mtight&quot;&gt;∗&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 수렴한다.&lt;/li&gt;&lt;li&gt;그러나 만약 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mo&gt;&amp;gt;&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\lambda &amp;gt; 0&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.73354em;vertical-align:-0.0391em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;λ&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;&amp;gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이면, loss에 기존의 weight L2 노름 제곱을 더해준다. 즉, 원래 weight가 크면 클수록 loss가 커지므로 penalty를 먹는다.&lt;ul&gt;&lt;li&gt;따라서 자연스럽게 학습은 weight를 최소화하는 방향으로 간다. 그래서 초록점쪽으로 이동하게 된다.&lt;/li&gt;&lt;li&gt;결과적으로 둘 사이의 어중간한 빨간 점에서 멈추게 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;그런데 Pruning은 weight를 직접적으로 잘라내는 것이기 때문에, regularization term에 직접적으로 영향을 미친다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:63.671875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;pruning_acc_loss&quot; title=&quot;pruning_acc_loss&quot; src=&quot;/static/809d89b069a52d85bf9aee16e0f47439/2bef9/pruning_acc_loss.png&quot; srcSet=&quot;/static/809d89b069a52d85bf9aee16e0f47439/6f3f2/pruning_acc_loss.png 256w,/static/809d89b069a52d85bf9aee16e0f47439/01e7c/pruning_acc_loss.png 512w,/static/809d89b069a52d85bf9aee16e0f47439/2bef9/pruning_acc_loss.png 1024w,/static/809d89b069a52d85bf9aee16e0f47439/71c1d/pruning_acc_loss.png 1536w,/static/809d89b069a52d85bf9aee16e0f47439/799f3/pruning_acc_loss.png 1642w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위의 그래프를 보면, L2 regularization은 retrain을 했을 때와 그렇지 않을때의 차이가 극적으로 나는 것을 확인할 수 있다. L2 regularization은 곧 weight가 0이지 않으면 페널티를 먹인다는 것과 같은 말이므로, weight의 크기를 줄여 영향력을 최소화하는 것이다.&lt;/p&gt;&lt;p&gt;이처럼 크기가 큰 weight를 줄여 overfitting을 막고 generalize하면, pruning 기법을 수행했을 때 90%나 되는 모델의 weight를 모두 잘라버려도 남은 weight들이 충분히 일반화되어있기 때문에 accuracy에 별 차이가 없어진다. 자르기전이나, 자른 후나 weight들의 형태가 고만고만하기 때문이다.&lt;/p&gt;&lt;h2 id=&quot;여러-pruning-기법&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%97%AC%EB%9F%AC-pruning-%EA%B8%B0%EB%B2%95&quot; aria-label=&quot;여러 pruning 기법 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;여러 Pruning 기법&lt;/h2&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:55.859375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;pruning_methods&quot; title=&quot;pruning_methods&quot; src=&quot;/static/c850b8419032d575c366ee6e91ed4c90/2bef9/pruning_methods.png&quot; srcSet=&quot;/static/c850b8419032d575c366ee6e91ed4c90/6f3f2/pruning_methods.png 256w,/static/c850b8419032d575c366ee6e91ed4c90/01e7c/pruning_methods.png 512w,/static/c850b8419032d575c366ee6e91ed4c90/2bef9/pruning_methods.png 1024w,/static/c850b8419032d575c366ee6e91ed4c90/71c1d/pruning_methods.png 1536w,/static/c850b8419032d575c366ee6e91ed4c90/a878e/pruning_methods.png 2048w,/static/c850b8419032d575c366ee6e91ed4c90/5b712/pruning_methods.png 2692w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;무엇을 잘라낼 것인가?(What to prune?)&lt;ul&gt;&lt;li&gt;&lt;strong&gt;unstructured&lt;/strong&gt; : 무작위로 개개의 weight들을 잘라내기&lt;/li&gt;&lt;li&gt;&lt;strong&gt;structured&lt;/strong&gt; : 단위를 잡아서 한번에 잘라내기&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;어떻게 잘라낼 것인가?(How to prune?)&lt;/li&gt;&lt;li&gt;언제 잘라낼 것인가?(When to prune?)&lt;/li&gt;&lt;li&gt;얼마나 자주 잘라낼 것인가?(How often?)&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:41.79687500000001%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAICAIAAAB2/0i6AAAACXBIWXMAABYlAAAWJQFJUiTwAAABIUlEQVQY03VQXUtDMQz1//8OQRFfRRARJ3MPsu1hygZuUzbBuS+9u7dt0iZtPXdjoA7DoaRpTnNyjnLOqhpEfAgAclRSrmPx3hv3r6bDxtuoMR3ejvvXL4Ob+eS+WnbLRZvL4RGaZsvVbLX+MmZdVRUR2JISzsmo2W4ed1unvYezTuvkqXP+OriYPl+W82b5cUfFY002RNZ7F9V5JlUjUpKzjNyzMguHFBQf5rjVhFO20JrsQqDNpiRb2A07R4GNJy9C7D9tZUWdRBuUUpT9RrvYkus+IlMZjErJe2asn3NIER6AgNxBnXjNv6Imk0hISfZq/iDsJO6RDyfXw1X/A8cIO2GqHJIxllXhmWH2MUJI6RwqAMzHFUW8LosCLv4kfwO8RcwwNBbnJQAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;iterative_pruning&quot; title=&quot;iterative_pruning&quot; src=&quot;/static/1d5ed7a8a8c318214a6d28bb4b087ca9/2bef9/iterative_pruning.png&quot; srcSet=&quot;/static/1d5ed7a8a8c318214a6d28bb4b087ca9/6f3f2/iterative_pruning.png 256w,/static/1d5ed7a8a8c318214a6d28bb4b087ca9/01e7c/iterative_pruning.png 512w,/static/1d5ed7a8a8c318214a6d28bb4b087ca9/2bef9/iterative_pruning.png 1024w,/static/1d5ed7a8a8c318214a6d28bb4b087ca9/71c1d/iterative_pruning.png 1536w,/static/1d5ed7a8a8c318214a6d28bb4b087ca9/a878e/iterative_pruning.png 2048w,/static/1d5ed7a8a8c318214a6d28bb4b087ca9/863d7/iterative_pruning.png 3520w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;Pruning은 한번에 수행되지 않고, 가지치기 이후 retraining(fine tuning이나 from scratch)과정을 몇번씩 거친다. 한번에 수행되면 weight들이 다 잘려나가 성능이 급격히 떨어진다. 그보다는, 여러번 반복하여 성능을 복원했다가 야금야금 pruning하는 방법이 주로 사용된다.&lt;/p&gt;&lt;h2 id=&quot;lottery-ticket-hypothesis&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#lottery-ticket-hypothesis&quot; aria-label=&quot;lottery ticket hypothesis permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Lottery ticket hypothesis&lt;/h2&gt;&lt;p&gt;2015년에 제안되었으며, research 트렌드를 주도했던 논문이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:62.890625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;lottery-ticket&quot; title=&quot;lottery-ticket&quot; src=&quot;/static/7ac53fc2a0bc5eb8e46a594b951738ff/2bef9/lottery-ticket.png&quot; srcSet=&quot;/static/7ac53fc2a0bc5eb8e46a594b951738ff/6f3f2/lottery-ticket.png 256w,/static/7ac53fc2a0bc5eb8e46a594b951738ff/01e7c/lottery-ticket.png 512w,/static/7ac53fc2a0bc5eb8e46a594b951738ff/2bef9/lottery-ticket.png 1024w,/static/7ac53fc2a0bc5eb8e46a594b951738ff/4352a/lottery-ticket.png 1364w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위의 이미지를 보면, pruning 전에 traning을 수행한 네트워크의 accuracy와, pruning 후에 training을 수행한 accuracy가 동일함을 알 수 있다. 즉, &lt;strong&gt;&lt;div&gt;기존의 네트워크에서 정말로 accuracy에 영향을 미치는 부분 네트워크는 따로있다는 것&lt;/div&gt;&lt;/strong&gt;이다. 이런 가설을 제시했던 것이 &lt;code&gt;lottery-ticket hypothesis&lt;/code&gt;이다.&lt;/p&gt;&lt;p&gt;subnetwork는 original 네트워크보다 파라미터도 적고, epoch을 적게 돌면서도 accuracy는 같거나 더 높은, &amp;#x27;로또&amp;#x27; 네트워크이다. 이 때 로또 네트워크는 반드시 original 네트워크의 random initializing을 따라가야 효과가 있는 subnetwork이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:37.5%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;pruning_techniques&quot; title=&quot;pruning_techniques&quot; src=&quot;/static/c011e2345352f90f633827c47579bc70/2bef9/pruning_techniques.png&quot; srcSet=&quot;/static/c011e2345352f90f633827c47579bc70/6f3f2/pruning_techniques.png 256w,/static/c011e2345352f90f633827c47579bc70/01e7c/pruning_techniques.png 512w,/static/c011e2345352f90f633827c47579bc70/2bef9/pruning_techniques.png 1024w,/static/c011e2345352f90f633827c47579bc70/71c1d/pruning_techniques.png 1536w,/static/c011e2345352f90f633827c47579bc70/a878e/pruning_techniques.png 2048w,/static/c011e2345352f90f633827c47579bc70/ca435/pruning_techniques.png 3730w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;Searching for Tickets&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;반복적으로 Magnitude(weight의 크기) 가지치기 - 가장 낮은 것부터 일정 부분을 잘라냄&lt;/li&gt;&lt;li&gt;random initialization한 네트워크를 가지고 pruning 하고, 다시 잘라낸 뒤의 네트워크를 초기 값(random initialization 했던 값)으로 다시 가지치기하고...&lt;/li&gt;&lt;li&gt;이걸 계속 반복하여 로또 티켓이 되는 노드들만 남긴 네트워크를 찾는 것.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Iterative Magnitude Pruning with Rewinding&lt;ul&gt;&lt;li&gt;반복적으로 Magnitude(weight의 크기) 가지치기 - 가장 낮은것부터 일정 부분을 잘라냄&lt;/li&gt;&lt;li&gt;Searching for Tickets와 다른점은, pruning 이후 재 학습시 매번 초기값이 아니라 어느정도 학습시켜놓은(iteration k 시점의) weight로 초기화한다. → 되감기하는 것 같아서 rewinding이라고 한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;</content:encoded></item><item><title><![CDATA[Computer Vision 08 - Multi-modal Learning]]></title><description><![CDATA[Multi-modal Learning by 오태현 교수님, BoostCamp AI Tech 7주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/44_multi-modal_learning/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/44_multi-modal_learning/</guid><pubDate>Fri, 12 Mar 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;multi-modal-learning&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#multi-modal-learning&quot; aria-label=&quot;multi modal learning permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Multi-modal learning&lt;/h1&gt;&lt;p&gt;기계 학습에 사용되는 데이터의 종류는 다양하다. 일반적으로 이 데이터 중 특정 형태의 데이터만을 모아 학습시키는 것이 기존의 unimodal learning이었지만, &lt;strong&gt;&lt;code&gt;multi-modal learning&lt;/code&gt;&lt;/strong&gt; 에서는 각기 다른 여러 종류의 데이터를 모두 사용하여 학습한다.&lt;/p&gt;&lt;p&gt;서로 다른 데이터 양식(modalities)을 사용할 때 문제점이 몇 가지 생긴다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;표현하는 방법이 다르다.&lt;ul&gt;&lt;li&gt;음성은 파형, 이미지는 픽셀별 matrix, 텍스트는 워드 임베딩 벡터 등...&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;서로 다른 형태에서 오는 정보의 양(feature spaces)도 unbalance하다.&lt;ul&gt;&lt;li&gt;어떤 텍스트를 주고 그에 맞는 이미지를 달라고 했을 때, 텍스트의 요구사항을 만족하는 이미지는 N개가 존재한다. 거꾸로, 그 이미지 중 하나를 텍스트로 바꿀수는 없다. 1:1 대응관계가 아니라 1:N이기 때문이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;여러 데이터를 넣더라도, 학습과정에서 모델이 사용하기 쉬운 데이터에 편향된 학습을 진행해 모델의 성능이 잘 나오지 않을 수 있다.&lt;ul&gt;&lt;li&gt;영상 데이터에서 소리를 듣지 않고 입모양으로도 말한다는 것을 catch할 수 있다면, 데이터를 모두 넣어줬더라도 소리 데이터를 비교적 적게 참조하고 영상 데이터의 학습 중 참조 비중이 커질것이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;그럼에도 불구하고 기존의 학습 한계점들을 해결할 수 있는 방법이기때문에 연구가 활발히 지속되고 있다.&lt;/p&gt;&lt;p&gt;multi-modal을 사용하는 task들의 패턴은 다음과 같다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Matching&lt;/code&gt;&lt;/strong&gt; : 서로 다른 modality의 data를 공통된 영역으로 가져가 비교하여 매칭한다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Translating&lt;/code&gt;&lt;/strong&gt; : 하나의 modality data를 다른 modality data로 translation한다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Referencing&lt;/code&gt;&lt;/strong&gt; : A modality에서 B modality로 출력하고싶을 때, B와 같은 형식의 B&amp;#x27;를 참고하여 출력한다.&lt;/li&gt;&lt;/ol&gt;&lt;h1 id=&quot;multi-modal-tasks--image--text&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#multi-modal-tasks--image--text&quot; aria-label=&quot;multi modal tasks  image  text permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Multi-modal tasks : Image &amp;amp; text&lt;/h1&gt;&lt;h2 id=&quot;matching&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#matching&quot; aria-label=&quot;matching permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Matching&lt;/h2&gt;&lt;p&gt;Visual data &amp;amp; text data를 matching하는 사례를 살펴보자.&lt;/p&gt;&lt;h3 id=&quot;text-embedding&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#text-embedding&quot; aria-label=&quot;text embedding permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Text embedding&lt;/h3&gt;&lt;p&gt;text 데이터는 word 단위로 dense vector로 임베딩한다. 임베딩된 dense representation들을 학습하면 일반화 성능이 생긴다.&lt;/p&gt;&lt;p&gt;임베딩 방식은 &lt;code&gt;word2vec(Skip-gram model)&lt;/code&gt;을 사용한다. &lt;/p&gt;&lt;p&gt;&lt;img src=&quot;https://www.researchgate.net/publication/322905432/figure/fig1/AS:614314310373461@1523475353979/The-architecture-of-Skip-gram-model-20.png&quot; alt=&quot;https://www.researchgate.net/publication/322905432/figure/fig1/AS:614314310373461@1523475353979/The-architecture-of-Skip-gram-model-20.png&quot;/&gt;&lt;/p&gt;&lt;ol&gt;&lt;li&gt;text 내 한 단어의 원-핫벡터를 한 row로 하는 입력을 넣는다.&lt;/li&gt;&lt;li&gt;W를 곱하여 얻은 hidden-layer의 벡터가 워드 임베딩 벡터가 된다.&lt;/li&gt;&lt;li&gt;특정 단어의 워드임베딩 벡터를 중심으로 하여, 주위의 단어들을 예측한다. 중심 단어의 워드 임베딩 벡터에 W&amp;#x27;를 곱하여서 만들어낸 많은 y들이 주위 단어들이다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;Skip-gram 모델은 학습 시 단어들간의 관계를 알기 위하여 N개의 이웃하는 단어들을 예측하도록 학습한다.&lt;/p&gt;&lt;h3 id=&quot;joint-embedding&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#joint-embedding&quot; aria-label=&quot;joint embedding permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Joint embedding&lt;/h3&gt;&lt;p&gt;Matching을 하기 위한 공통의 임베딩 벡터들을 학습하는 방식이다. Image Tagging의 사례를 생각해보자. 이미지가 주어지면 그 이미지를 표현하는 여러 단어들을 작성하도록 할 수도 있고, 여러 단어들을 주고 이에 해당하는 이미지를 찾을수도 있을 것이다.&lt;/p&gt;&lt;p&gt;각 pre-trained unimodal model들을 가져와서 각 형식의 데이터를 feature vector 형태로 표시한다. 이때, 두 모델의 output dimension은 동일한 크기로 고정한다(같은 space를 공유하도록). 이후, Joint embedding 부분에서 만약 두 output이 다르면(이미지와 단어가 잘 맞지않으면) 큰 distance를 주고, 그렇지 않으면 작은 distance를 주어 학습시킨다. 이런 방식으로 크고 작은 distance를 부여하여 학습시키는 방법을 &lt;code&gt;metric learning&lt;/code&gt;이라고 한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;신기하게도, 이렇게 학습을 시킨 joint embedding은 visual 데이터와 text 데이터 사이의 multi-modal analogy relationship까지 모두 가지고 있다. 개 이미지에서 개(단어)를 빼고, 고양이(단어)를 추가하면 고양이 이미지가 나온다.&lt;/li&gt;&lt;li&gt;이처럼 이미지와 단어의 관계까지 학습하고 있다보니, 요리 이미지를 넣었을 때 레시피(재료+요리방법)를 주거나 레시피를 넣었을 때 해당 레시피로 만든 요리 이미지를 주는 모델을 만드는 등의 응용도 가능하다. 재료와 요리 방법을 각각 다른 워드 임베딩 모델로 임베딩하여 얻어낸 feature map을 concat하고, 이미지 feature map과 dim을 맞추어 joint-embedding하면 된다.&lt;ul&gt;&lt;li&gt;cosine similarity loss로 joint embedding을 학습하고, semantic regularization loss를 사용하여 high-level semantic(예를 들어 레시피와 이미지가 정확히 맞진 않더라도, &amp;#x27;튀김요리&amp;#x27;라는 카테고리 정도는 맞으면 좋겠다는 식의 가이드)을 학습한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;translating--cross-modal-translation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#translating--cross-modal-translation&quot; aria-label=&quot;translating  cross modal translation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Translating : Cross modal translation&lt;/h2&gt;&lt;p&gt;대표적인 translation의 사례는 Image captioning이 있다. Image를 문장으로 바꾸거나(CNN), 문장을 이미지로 바꾸어(RNN) caption 관계를 만들 수 있다. 결국은 CNN과 RNN을 잘 합쳐야한다는 말이다.&lt;/p&gt;&lt;h3 id=&quot;show-and-tell&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#show-and-tell&quot; aria-label=&quot;show and tell permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Show and tell&lt;/h3&gt;&lt;p&gt;ImageNet에서 pre-train된 CNN 모델을 인코더로 사용하고, 디코더로 LSTM 모듈을 사용한다. 하나의 fixed dimensional 벡터에서 이미지 캡션 전체를 한번에 prediction 한다.&lt;/p&gt;&lt;h3 id=&quot;show-attend-tell&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#show-attend-tell&quot; aria-label=&quot;show attend tell permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Show, attend, tell&lt;/h3&gt;&lt;p&gt;캡셔닝을 할 때, 좀 더 주목해야하는 단어들이 있기 마련이다. 예를 들어 공 놀이를 하는 소년의 사진이라면, &amp;#x27;공&amp;#x27;과 &amp;#x27;소년&amp;#x27;의 이미지에 주목하여 이미지를 생성해야한다.  &lt;em&gt;Show, attend, tell&lt;/em&gt; 모델은 attention을 사용하여 좀 더 정확한 의미로 캡셔닝을 수행하도록 한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:40.625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;show-attend-tell&quot; title=&quot;show-attend-tell&quot; src=&quot;/static/aaab8e81cfff55fb0abab03d79251177/2bef9/show-attend-tell.png&quot; srcSet=&quot;/static/aaab8e81cfff55fb0abab03d79251177/6f3f2/show-attend-tell.png 256w,/static/aaab8e81cfff55fb0abab03d79251177/01e7c/show-attend-tell.png 512w,/static/aaab8e81cfff55fb0abab03d79251177/2bef9/show-attend-tell.png 1024w,/static/aaab8e81cfff55fb0abab03d79251177/71c1d/show-attend-tell.png 1536w,/static/aaab8e81cfff55fb0abab03d79251177/a878e/show-attend-tell.png 2048w,/static/aaab8e81cfff55fb0abab03d79251177/da5ba/show-attend-tell.png 3238w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;CNN이 fixed vector을 출력하지 않고, 공간 정보를 유지하고있는 14x14 feature map을 출력하여 RNN에 넣어준다. 단어 생성시마다 이 feature map을 referencing한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:70.703125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;soft-attention&quot; title=&quot;soft-attention&quot; src=&quot;/static/9057ed7dcc9049b529652868511032bb/2bef9/soft-attention.png&quot; srcSet=&quot;/static/9057ed7dcc9049b529652868511032bb/6f3f2/soft-attention.png 256w,/static/9057ed7dcc9049b529652868511032bb/01e7c/soft-attention.png 512w,/static/9057ed7dcc9049b529652868511032bb/2bef9/soft-attention.png 1024w,/static/9057ed7dcc9049b529652868511032bb/71c1d/soft-attention.png 1536w,/static/9057ed7dcc9049b529652868511032bb/df88b/soft-attention.png 1906w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;spartial한 feature map이 RNN으로 들어오면, condition을 줘서 어디를 reference해야하는지 heat map을 만들고, 이 &lt;strong&gt;heat map과 feature map을 inner product한 weighted sum z 벡터&lt;/strong&gt;를 만든다.&lt;ul&gt;&lt;li&gt;attention weight는 probaility가 되고, featrue map은 weight가 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;text-to-image-by-generative-model&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#text-to-image-by-generative-model&quot; aria-label=&quot;text to image by generative model permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Text-to-image by generative model&lt;/h3&gt;&lt;p&gt;이미지에서 캡셔닝을 하는것이 아니라, 텍스트에서 이에 맞는 이미지를 찾아내는 것은 어떨까? 이 경우, 하나의 텍스트에 대응되는 이미지는 여러 장이 존재한다. 이처럼 1:N의 관계로 출력물을 내야하는 경우 자연스럽게 Generative Model을 사용해볼 수 있다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:28.125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;text-to-image-by-generative-model&quot; title=&quot;text-to-image-by-generative-model&quot; src=&quot;/static/586f20f74c96b773471976f9ecf6bae6/2bef9/text-to-image-by-generative-model.png&quot; srcSet=&quot;/static/586f20f74c96b773471976f9ecf6bae6/6f3f2/text-to-image-by-generative-model.png 256w,/static/586f20f74c96b773471976f9ecf6bae6/01e7c/text-to-image-by-generative-model.png 512w,/static/586f20f74c96b773471976f9ecf6bae6/2bef9/text-to-image-by-generative-model.png 1024w,/static/586f20f74c96b773471976f9ecf6bae6/71c1d/text-to-image-by-generative-model.png 1536w,/static/586f20f74c96b773471976f9ecf6bae6/a878e/text-to-image-by-generative-model.png 2048w,/static/586f20f74c96b773471976f9ecf6bae6/3b25c/text-to-image-by-generative-model.png 3440w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;Generative Model(인코더)을 학습할 때, fixed dimensional vector로 만들어주는 네트워크에 Gaussian random code를 붙여서 diverser한 output을 만들 수 있도록 해준다. Condition, Input, Sentence 정보가 모두 Generator에 들어가는 모델이다.&lt;/p&gt;&lt;p&gt;Discriminator Model(디코더)는 Generative model이 생성한 이미지를 받아서, row dimensional spatial vector를 뽑아내고, Generative model이 사용했던 sentence를 받아 이미지와의 정합성을 판단하도록 학습한다.&lt;/p&gt;&lt;h2 id=&quot;referencing--cross-modal-reasoning&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#referencing--cross-modal-reasoning&quot; aria-label=&quot;referencing  cross modal reasoning permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Referencing : Cross modal reasoning&lt;/h2&gt;&lt;h3 id=&quot;visual-question-answering&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#visual-question-answering&quot; aria-label=&quot;visual question answering permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Visual question answering&lt;/h3&gt;&lt;p&gt;영상과 질문이 주어지면 그에 맞는 대답을 하는 task이다. 텍스트는 RNN으로, 이미지는 Pre-trained NN으로, 각각의 fixed dimensional vector로 만들어준다. 이후 두 벡터들을 Point-wise multiplication으로 두 임베딩 feature가 서로 interaction 할 수 있게 만들어준다. 일종의 joint embedding space라고 볼 수 있다. 이 과정을 end-to-end로 학습시킨다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:38.67187499999999%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;visual-question-answering&quot; title=&quot;visual-question-answering&quot; src=&quot;/static/0976ccbfba5574faa471ea67a71fb9c4/2bef9/visual-question-answering.png&quot; srcSet=&quot;/static/0976ccbfba5574faa471ea67a71fb9c4/6f3f2/visual-question-answering.png 256w,/static/0976ccbfba5574faa471ea67a71fb9c4/01e7c/visual-question-answering.png 512w,/static/0976ccbfba5574faa471ea67a71fb9c4/2bef9/visual-question-answering.png 1024w,/static/0976ccbfba5574faa471ea67a71fb9c4/71c1d/visual-question-answering.png 1536w,/static/0976ccbfba5574faa471ea67a71fb9c4/a878e/visual-question-answering.png 2048w,/static/0976ccbfba5574faa471ea67a71fb9c4/61410/visual-question-answering.png 3224w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;이외에도 어텐션 메커니즘을 사용하여 cross modal reasoning을 수행할수도 있다. 각각의 modal을 referencing하며 해결하는 형태이다.&lt;/p&gt;&lt;h1 id=&quot;multi-modal-tasks--audio&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#multi-modal-tasks--audio&quot; aria-label=&quot;multi modal tasks  audio permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Multi-modal tasks : Audio&lt;/h1&gt;&lt;h2 id=&quot;sound-representation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#sound-representation&quot; aria-label=&quot;sound representation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Sound Representation&lt;/h2&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:29.6875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;sound-representation&quot; title=&quot;sound-representation&quot; src=&quot;/static/7da33f3dc2780085e1a0d4476f499c43/2bef9/sound-representation.png&quot; srcSet=&quot;/static/7da33f3dc2780085e1a0d4476f499c43/6f3f2/sound-representation.png 256w,/static/7da33f3dc2780085e1a0d4476f499c43/01e7c/sound-representation.png 512w,/static/7da33f3dc2780085e1a0d4476f499c43/2bef9/sound-representation.png 1024w,/static/7da33f3dc2780085e1a0d4476f499c43/71c1d/sound-representation.png 1536w,/static/7da33f3dc2780085e1a0d4476f499c43/a878e/sound-representation.png 2048w,/static/7da33f3dc2780085e1a0d4476f499c43/88b57/sound-representation.png 3582w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;Sound는 원래 wave 데이터 형태로 존재하는데, ML/DL에서 사용할 때는 Power spectrum이나 Spectogram, MFCC같은 Acoustic feature들로 변환해서 사용한다. 이런 변환 방식에 대해 알아보자.&lt;/p&gt;&lt;h3 id=&quot;short-time-fourier-transformstft-&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#short-time-fourier-transformstft-&quot; aria-label=&quot;short time fourier transformstft  permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Short-time Fourier transform(STFT )&lt;/h3&gt;&lt;p&gt;시그널 프로세싱에서 가장 많이 사용되는 푸리에 변환을 변형시킨 방식이다. 시간 축(t) 전체를 그냥 주파수형태로 모두 옮겨버리면, 시간에 따른 변화를 파악할 수 없다. 그래서 제안된 것이 STFT이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:31.25%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA40lEQVQY032QSW7DMAxFff8r9AZd9hBZtItk00VRoHCcwbGrwY0GUyQllXEAI+iihBYf0HufkJr67wDAt1IJqZSac3k4WW6bBHPGJKmUkgVZAjOXJd985Jnyn1KRBWjM6aI/25rSSmtthnEUn7iWXNVPtAGX0ipEeZT3bf/0vHl9P7196Y+LJy7BB2Otc16banTdj65V/i6vc9/UDIfzdrPzU7Ser4FCIABOyLfuheu0e9keOxX6KRqftAcPGGcgoqYQ1gST1tYaaZQfEgsR11dIAKQ54bkfusPR+cCcr87FGH8BO89bjtVyQGsAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;stft&quot; title=&quot;stft&quot; src=&quot;/static/791e3c94b2cad2cecd322697805402e1/2bef9/stft.png&quot; srcSet=&quot;/static/791e3c94b2cad2cecd322697805402e1/6f3f2/stft.png 256w,/static/791e3c94b2cad2cecd322697805402e1/01e7c/stft.png 512w,/static/791e3c94b2cad2cecd322697805402e1/2bef9/stft.png 1024w,/static/791e3c94b2cad2cecd322697805402e1/71c1d/stft.png 1536w,/static/791e3c94b2cad2cecd322697805402e1/a878e/stft.png 2048w,/static/791e3c94b2cad2cecd322697805402e1/ea7d0/stft.png 3620w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위의 그림처럼 슬라이딩 윈도우로 시간축에 따른 표시를 하되, 튀는 부분들을 자르고 가운데 부분들을 강조하는 형태로 windowing을 시킨다(Hamming window).  자르는 크기와 시간등은 하이퍼파라미터이다.&lt;/p&gt;&lt;p&gt;다만, 윈도우를 완전히 겹치는게 아니라 너무 dense하지 않는 선에서 적당히만 overlap 시켜가며 찍어 spectogram으로 만들어준다.&lt;/p&gt;&lt;p&gt;푸리에 변환을 왜 할까? 시간 축의 input signal이 있을 때, 푸리에 변환을 통해 주파수의 삼각함수의 성분을 분해할 수 있기 때문이다. Power spectrum에서 각 feature 별로 frequency를 파악할 수 있게 된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;ex - 남성의 목소리는 저주파에 있고, 여성의 목소리는 상대적으로 고주파에 있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;spectogram은 시간축에서 spectrum들이 쌓인 것인데, 특정 성분이 많을 경우 밝은 흰색으로 표시된다. 이를 통해 특정 시간에 어떤 성분들로 이루어져있는지를 확인할 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;요새는 딥러닝 신경망을 사용할 때, spectogram을 바로 사용하는 경우도 많고, dimension을 조금 낮춰서 melspectogram을 사용하는 경우도 있다. MFCC도 검색해보자.&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;joint-embedding-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#joint-embedding-1&quot; aria-label=&quot;joint embedding 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Joint embedding&lt;/h2&gt;&lt;h3 id=&quot;sound-tagging&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#sound-tagging&quot; aria-label=&quot;sound tagging permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Sound tagging&lt;/h3&gt;&lt;p&gt;현재 소리가 어느 장소(ex - 해변, 교실 등)에서 들리는 소리인지를 파악하는 task이다.&lt;/p&gt;&lt;p&gt;SoundNet은 입력 이미지를 받아 두 개의 Pretrained ImageNet을 통해 Object distribution과 Scene(Place) Distribution을 출력한다. 또한, raw audio를 wave 데이터 형태로 추출해 CNN 구조에 넣어주어, 두개의 head를 최종 층에 두고 각각의 distribution을 사용하여 Place 구성을 따라하고, Object recognition을 수행하도록 한다(KL divergence minimize).&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때 spectogram 대신 waveform을 쓴 것은 특별한 사유가 있는건 아니고, 이 논문이 나왔을 무렵 아직 spectogram 같은 변환 데이터가 자주 사용되지 않았기 때문이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 방식은 기학습된 ImageNet을 fix해두고, 그 분포를 따라가도록 조그마한 CNN 모델을 학습시키는 것이기 때문에 Teacher-student 모델이라고 할 수 있다.&lt;/p&gt;&lt;p&gt;원하는 target task에 pre-trained 모델을 응용할수도 있을텐데, 이 경우 CNN 모델의 Pool5 feature를 뽑아서 classifier를 올려 그 부분만 target task에 대해 학습하도록 한다. 이렇게 하는 이유는 Teacher 모델의 object/scene에 너무 특화되지 않도록, generalization하기 위해서이다.&lt;/p&gt;&lt;h2 id=&quot;cross-modal-translation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#cross-modal-translation&quot; aria-label=&quot;cross modal translation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Cross modal translation&lt;/h2&gt;&lt;h3 id=&quot;speech2face&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#speech2face&quot; aria-label=&quot;speech2face permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Speech2Face&lt;/h3&gt;&lt;p&gt;음성을 듣고 그 사람의 얼굴을 상상해내는 네트워크이다. Spectogram이 input으로 들어가고, fixed dimensional vector가 들어오면 Face Decoder에 들어가서 얼굴을 생성해준다.&lt;/p&gt;&lt;p&gt;Speech2Face 모델은 여러 모델을 잘 조합시킨 Module Network라고 할 수 있다. Face Recognition 모델로 VGG-Face Model, Face feature를 정규화된 이미지로 reconstruction하는 Face Decoder, Spectogram을 Voice로 바꾸어주는 Voice Encoder등으로 구성되어있다. Face 벡터가 Voice를 따라하도록 된 구조이다.&lt;/p&gt;&lt;h3 id=&quot;image-to-speech-synthesis&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#image-to-speech-synthesis&quot; aria-label=&quot;image to speech synthesis permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Image-to-speech synthesis&lt;/h3&gt;&lt;p&gt;이미지를 넣으면 음성을 만들어주는 방식이다.&lt;/p&gt;&lt;p&gt;이미지에서 단어를 바로 추출하는 것은 아니고, sub-word라고 불리는 유닛(토큰) 형태로 변환하는 Image-to-Unit 모델과, 유닛에서 음성으로 변환하는 Unit-to-Speech 모델 두개의 조합으로 이루어져있다.&lt;/p&gt;&lt;h3 id=&quot;sound-source-localization&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#sound-source-localization&quot; aria-label=&quot;sound source localization permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Sound source localization&lt;/h3&gt;&lt;p&gt;소리를 input으로 넣고 이미지를 주었을 때, 이미지의 어느 부분에서 이 소리가 나는지 위치를 짚어주는 task이다. 소리의 context를 파악하고, 이 소리의 semantic한 의미와 이미지의 semantic context를 추출하여 비교할 수 있어야 한다.&lt;/p&gt;&lt;p&gt;Visual Network에서 공간정보를 얻고, Audio Network의 출력을 Attention Network에서 내적하여 관계를 규명한다.&lt;/p&gt;&lt;p&gt;Ground truth가 있다면 supervised learning을 수행할 수 있다. 그러나 일반적으로 video는 sound를 동반하므로, 그냥 video을 annotation처럼 활용하여 unsupervised learning도 가능하다. Audio network의 출력물과 attention network의 출력물 visual feature을 metric learning한다.&lt;/p&gt;&lt;h2 id=&quot;cross-modal-reasoning&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#cross-modal-reasoning&quot; aria-label=&quot;cross modal reasoning permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Cross modal reasoning&lt;/h2&gt;&lt;h3 id=&quot;looking-to-listen-at-the-cocktail-party&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#looking-to-listen-at-the-cocktail-party&quot; aria-label=&quot;looking to listen at the cocktail party permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Looking to listen at the cocktail party&lt;/h3&gt;&lt;p&gt;하나의 비디오에서 여러 사람이 말하는 것을 분리하는 task이다. 이 경우 train data를 찾기 힘들다. 마땅한 pairwise data가 잘 없기 때문이다. 그래서 조금 더 smart한 방법으로, 한 사람이 말하는 비디오 여러 개를 합성하여 train dataset을 만든다.&lt;/p&gt;&lt;h3 id=&quot;lip-movements-generation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#lip-movements-generation&quot; aria-label=&quot;lip movements generation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Lip movements generation&lt;/h3&gt;&lt;p&gt;sound source로부터 특정 사람의 목소리, 얼굴, 입모양을 생성해내는 task이다.&lt;/p&gt;&lt;p&gt;9강 실습 미완&lt;/p&gt;&lt;hr/&gt;&lt;h3 id=&quot;reference&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#reference&quot; aria-label=&quot;reference permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Reference&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://www.researchgate.net/figure/The-architecture-of-Skip-gram-model-20_fig1_322905432&quot;&gt;The architecture of Skip-gram Model&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[Computer Vision 06 - Instance & Panoptic segmentation]]></title><description><![CDATA[Instance segmentation and Panoptic segmentation by 오태현 교수님, BoostCamp AI Tech 7주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/42_instance_and_panoptic_segmentation/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/42_instance_and_panoptic_segmentation/</guid><pubDate>Thu, 11 Mar 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;Instance,Panoptic segmentation&lt;/p&gt;&lt;h1 id=&quot;instance-segmentation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#instance-segmentation&quot; aria-label=&quot;instance segmentation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Instance segmentation&lt;/h1&gt;&lt;p&gt;  Semantic segmentation과의 Instance segementation이 다른 점은, 전자는 클래스만 구별하고 개체를 구별하지 않지만, 후자는 개체(Instance)까지도 구별한다는 점이다. 즉, 기존의 Semantic segmentation에서  distinguish instances라는 새로운 목표가 추가된 task이다.&lt;/p&gt;&lt;h2 id=&quot;mask-r-cnn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#mask-r-cnn&quot; aria-label=&quot;mask r cnn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Mask R-CNN&lt;/h2&gt;&lt;p&gt;  Faster R-CNN이 개량된 형태로, 큰 줄기는 비슷하다. Two-stage detector이다. 아래는 차이점들이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;RoI Pooling(정수 feature) → RoIAlign(소숫점 이하의 feature까지 지원)&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Interpolation을 통해 소수점 픽셀 레벨의 pooling을 지원하며, 이를 통해 더 정교한 feature를 뽑아 뒷단의 성능까지 향상시켰다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;Branch [class classification, box regression] → Branch [[class classification, box regression] + &lt;strong&gt;mask branch&lt;/strong&gt;]&lt;/p&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;text&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-text&quot;&gt;TEXT&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-text&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;- Mask Branch는 기존의 res를 upsampling하여 80개의 채널에 대해 binary mask로 prediction하는 구조이다.(Mask FCN predictor)&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;- 이 때 어떤 mask를 사용할 것인지는 class classification 결과에 따라 다르다.&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;h2 id=&quot;yolactyou-only-look-at-coefficients&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#yolactyou-only-look-at-coefficients&quot; aria-label=&quot;yolactyou only look at coefficients permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;YOLACT(You Only Look At CoefficienTs)&lt;/h2&gt;&lt;p&gt;Real-time으로 semantic segmentation이 가능한 Single-stage detector이다. &lt;/p&gt;&lt;p&gt;&lt;img src=&quot;https://miro.medium.com/max/1400/0*dFkJ_eVVZ3lxcw4U.png&quot; alt=&quot;https://miro.medium.com/max/1400/0*dFkJ_eVVZ3lxcw4U.png&quot;/&gt;&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;em&gt;FPN&lt;/em&gt;&lt;/strong&gt; : Feature pyramid를 통해 고해상도의 Feature map을 얻는 Network&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;em&gt;Protonet&lt;/em&gt;&lt;/strong&gt; : Mask는 아니지만, Mask를 합성해 낼 수 있는 base가 되는 Prototypes, 즉 여러 물체의 soft segmentation component들을 생성한다. span 가능한 basis라고 볼 수 있다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;em&gt;Prediction Head&lt;/em&gt;&lt;/strong&gt; : Protonet에서 나온 Prototypes를 적절하게 합성하기 위한 계수, Mask Coefficient를 출력한다.&lt;/p&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;text&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-text&quot;&gt;TEXT&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-text&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;- 이후 prototype들과 곱해져 적절한 detection response map(우상단)을 얻게 된다.&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;Mask를 효율적으로 생성하기 위해 Mask R-CNN처럼 Prototype 개수를 80개씩 만들지 않고 훨씬더 적은 개수를 설정하는 대신에, 그 때 그 때 coefficient와의 선형결합으로 찾아냄으로써 메모리 부담을 경감시켰다.&lt;/p&gt;&lt;h2 id=&quot;yolactedge&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#yolactedge&quot; aria-label=&quot;yolactedge permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;YolactEdge&lt;/h2&gt;&lt;p&gt;YOLOACT의 방식을 Edge computing에 접목하기 위해 계산량을 획기적으로 줄인 모델이다.&lt;/p&gt;&lt;p&gt;&lt;img src=&quot;https://programmersought.com/images/234/c25e6153b014fd65973238f43f6dadd2.png&quot; alt=&quot;https://programmersought.com/images/234/c25e6153b014fd65973238f43f6dadd2.png&quot;/&gt;&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;기존의 frame 중 key frame의 feature를 다음 frame에 전달(transform)하여 계산량을 크게 줄였다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;성능은 비슷하면서도 연산량이 많이 줄었기 때문에 edge device에서의 사용이 가능해졌다.&lt;/p&gt;&lt;p&gt;그러나 아직까지는 마스크가 떨린다던데, 깜빡인다던지 하는 문제가 있다.&lt;/p&gt;&lt;h1 id=&quot;panoptic-segmentation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#panoptic-segmentation&quot; aria-label=&quot;panoptic segmentation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Panoptic segmentation&lt;/h1&gt;&lt;p&gt;Instance segmentation은 배경에 관심이 없고, 그 위의 물체(객체)를 인스턴스별로 검출하는 데에 탁월했다. 배경을 검출하려면 차라리 Instance segmentation보다는, 개체를 구별할 수 없지만 semantic segmentation이 더 나았다. 이 두가지의 문제점들을 상호 보완하고 동시에 해결하려는 시도로 &lt;strong&gt;&lt;code&gt;panoptic segmentation&lt;/code&gt;&lt;/strong&gt;이 나왔다.(&lt;em&gt;Stuff + Instances of Things&lt;/em&gt;)&lt;/p&gt;&lt;h2 id=&quot;upsnet&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#upsnet&quot; aria-label=&quot;upsnet permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;UPSNet&lt;/h2&gt;&lt;p&gt;그야말로 Semantic segmentation과 Instance segmentation을 합친 형태이다.&lt;/p&gt;&lt;p&gt;&lt;img src=&quot;https://miro.medium.com/max/3070/1*k87beyGLIN3xsIG5OlxubQ.png&quot; alt=&quot;https://miro.medium.com/max/3070/1*k87beyGLIN3xsIG5OlxubQ.png&quot;/&gt;&lt;/p&gt;&lt;/li&gt;&lt;/ul&gt;&lt;ol&gt;&lt;li&gt;&lt;p&gt;Backbone Network는 FPN으로 고해상도 feature map을 뽑는다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;Semantic Head와 Instance Head가 각각의 역할을 수행한다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;마지막에, Panoptic head가 이 결과들을 융합하여 하나의 Panoptic logit으로 나타낸다.&lt;/p&gt;&lt;h3 id=&quot;architecture-of-the-panoptic-segmentation-head&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#architecture-of-the-panoptic-segmentation-head&quot; aria-label=&quot;architecture of the panoptic segmentation head permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Architecture of the panoptic segmentation head&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:44.140625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;upsnet-architecture&quot; title=&quot;upsnet-architecture&quot; src=&quot;/static/c82a286887adec897b06a8c6e157445a/2bef9/upsnet-architecture.png&quot; srcSet=&quot;/static/c82a286887adec897b06a8c6e157445a/6f3f2/upsnet-architecture.png 256w,/static/c82a286887adec897b06a8c6e157445a/01e7c/upsnet-architecture.png 512w,/static/c82a286887adec897b06a8c6e157445a/2bef9/upsnet-architecture.png 1024w,/static/c82a286887adec897b06a8c6e157445a/71c1d/upsnet-architecture.png 1536w,/static/c82a286887adec897b06a8c6e157445a/a878e/upsnet-architecture.png 2048w,/static/c82a286887adec897b06a8c6e157445a/8c565/upsnet-architecture.png 3346w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;/li&gt;&lt;/ol&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;Instance Head&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;Y&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Y_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;Y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.22222em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 개체 구별 Mask&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;Semantic Head&lt;/p&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;text&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-text&quot;&gt;TEXT&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-text&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;- stuff(background) : 그대로 최종 출력(Panoptic logit)에 들어간다.&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;- 물체(thing) : Instance head의 Instance response와 더해져서 최종 출력에 추가된다. 전체 출력에서 어느 위치에 해당 물체가 검출되어야하는지 찾아준다.&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;    - 어느 클래스로도 분류되지 않은 물체들은 Unknown class로 모두 합쳐서 1개의 층으로 최종 출력에 추가한다.&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;h2 id=&quot;vpsnetfor-video&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#vpsnetfor-video&quot; aria-label=&quot;vpsnetfor video permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;VPSNet(for video)&lt;/h2&gt;&lt;p&gt;단순한 이미지라면 한번의 구별로 끝이나겠지만, 비디오영상이라면 시시각각으로 움직이는 객체들이 매 순간 동일한 객체로 검출되어야할 것이다. VPSNet은 이런 문제를 각 시간별로 모든 픽셀의 대응 관계, 즉 motion을 catch하여 활용하는 Network이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:38.67187499999999%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;vpsnet-architecture&quot; title=&quot;vpsnet-architecture&quot; src=&quot;/static/9c92e9c5c77da6c28f7ff842b0344830/2bef9/vpsnet-architecture.png&quot; srcSet=&quot;/static/9c92e9c5c77da6c28f7ff842b0344830/6f3f2/vpsnet-architecture.png 256w,/static/9c92e9c5c77da6c28f7ff842b0344830/01e7c/vpsnet-architecture.png 512w,/static/9c92e9c5c77da6c28f7ff842b0344830/2bef9/vpsnet-architecture.png 1024w,/static/9c92e9c5c77da6c28f7ff842b0344830/71c1d/vpsnet-architecture.png 1536w,/static/9c92e9c5c77da6c28f7ff842b0344830/a878e/vpsnet-architecture.png 2048w,/static/9c92e9c5c77da6c28f7ff842b0344830/28fe3/vpsnet-architecture.png 3446w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;/li&gt;&lt;/ul&gt;&lt;ol&gt;&lt;li&gt;&lt;p&gt;이전 시점의 frame(reference frame) feature map을 현재 시점 frame(target frame)에 align하여 검출 률을 높이고, 시간에 따라 smooth한 검출률을 만든다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;여기서 Align한다는 것은 픽셀 단위로 Fusion한다는 의미이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;기존 RoI들과 새로 만들어진 현재의 RoI들을 tracking하여 매치함으로써 어떤 연관이 있는지 ML 학습한다.(Track Head) → 같은 물체는 같은 id를 가질 수 있게 된다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;각 결과들을 모두 합쳐 최종 출력 panoptic map을 형성한다.&lt;/p&gt;&lt;h1 id=&quot;landmark-localization&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#landmark-localization&quot; aria-label=&quot;landmark localization permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Landmark localization&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Landmark localization(Keypoint estimation)&lt;/code&gt;&lt;/strong&gt;은 키포인트를 정의하고 추정/tracking하는 데에 사용된다.&lt;/p&gt;&lt;/li&gt;&lt;/ol&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;얼굴이나 사람의 포즈를 tracking&lt;/p&gt;&lt;h2 id=&quot;coordinate-regression-vs-heatmap-classification&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#coordinate-regression-vs-heatmap-classification&quot; aria-label=&quot;coordinate regression vs heatmap classification permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Coordinate regression vs Heatmap classification&lt;/h2&gt;&lt;p&gt;Key points를 찾기 위해 사용할 수 있는 방법&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;Coordinate regression : 가장 간단하게 생각할 수 있는 방법으로, box regression처럼 각 포인트의 (x,y)를 사용한다. landmark가 N개라면 2N개가 나올 것이다. 그러나 이 방식은 대체로 부정확하고 일반화가 힘들다는 문제가 있었다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;code&gt;Heatmap classification&lt;/code&gt; : semantic segmentation처럼 각 채널이 각각의 keypoint를 담당하고, 각 keypoint가 일종의 class로 분류되어서 각 픽셀별로 classification하는 방법이다. 다만 훨씬 좋은 성능을 냈지만, 연산량이 너무 많았다(모든 픽셀에 대해 수행해야하기때문에)&lt;/p&gt;&lt;h3 id=&quot;landmark-location-to-gaussian-heatmap&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#landmark-location-to-gaussian-heatmap&quot; aria-label=&quot;landmark location to gaussian heatmap permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Landmark location to Gaussian heatmap&lt;/h3&gt;&lt;p&gt;Heatmap classification에서 x,y 위치(픽셀)가 주어졌을 때, 이걸 히트맵으로 변환하는 방법은 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;G_\sigma(x,y) = exp\bigg(-\frac{(x-x_c)^2+(y-y_c)^2}{2\sigma^2}\bigg)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.40003em;vertical-align:-0.95003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.441138em;vertical-align:-0.95003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.491108em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.740108em&quot;&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(x_c,y_c)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : center location&lt;/p&gt;&lt;h2 id=&quot;hourglass-network&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#hourglass-network&quot; aria-label=&quot;hourglass network permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Hourglass network&lt;/h2&gt;&lt;p&gt;Heatmap classification을 적절하게 활용한 네트워크구조로 &lt;code&gt;Hourglass Network&lt;/code&gt;가 있다.&lt;/p&gt;&lt;p&gt;&lt;img src=&quot;https://curt-park.github.io/images/stacked_hourglass_networks/fig1.png&quot; alt=&quot;https://curt-park.github.io/images/stacked_hourglass_networks/fig1.png&quot;/&gt;&lt;/p&gt;&lt;p&gt;U-Net과 비슷한 구조로, 모래시계같은 형태가 여러 겹 쌓였다고해서 &lt;strong&gt;Stacked Hourglass module&lt;/strong&gt;이라고도 부른다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;Hourglass 구조가 만들어진 이유&lt;/strong&gt;&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;영상 전체를 작게 만들어 receptive field를 키움으로써 전반적인 context를 보고 landmark를 파악한다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;low-level feature에서 skip-connection을 수행해서 정확한 landmark의 위치를 파악한다.&lt;/p&gt;&lt;p&gt;&lt;img src=&quot;https://curt-park.github.io/images/stacked_hourglass_networks/fig3.png&quot; alt=&quot;https://curt-park.github.io/images/stacked_hourglass_networks/fig3.png&quot;/&gt;&lt;/p&gt;&lt;p&gt;U-Net과 비슷하지만, skip connection이 &lt;strong&gt;concat이 아니라 Add&lt;/strong&gt;라는 것이다. 따라서 매 skip connection마다 dimension이 늘지 않는다. 대신, 단순히 skip하는 것이 아니라 또다른 conv layer를 통과하여 매핑해준다. 이런 점에서는 U-Net보다 FPN에 조금 더 가까운 구조라고 볼수도 있다(Add, Conv)&lt;/p&gt;&lt;h2 id=&quot;더-나아가기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%8D%94-%EB%82%98%EC%95%84%EA%B0%80%EA%B8%B0&quot; aria-label=&quot;더 나아가기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;더 나아가기&lt;/h2&gt;&lt;h3 id=&quot;densepose&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#densepose&quot; aria-label=&quot;densepose permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;DensePose&lt;/h3&gt;&lt;p&gt;신체의 모든 부위에 dense한 landmark를 찾는것 → 곧 3D surface를 얻는 것과 같다.&lt;/p&gt;&lt;p&gt;&lt;code&gt;UV map&lt;/code&gt;은 이러한 3D surface를 2D로 펼쳐서(flatten) 이미지 형태로 표현한 좌표 표기법이다. 즉, UV map의 2D 좌표 하나는 특정 3D 위치에 1:1로 매칭된다. 사실, UV map은 3D 모델 위에 texture map을 입히기 위해서 고안된 것이다. 이러한 좌표의 특성을 DensePose에서 차용하여 사용하고 있다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:43.359375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;densepose&quot; title=&quot;densepose&quot; src=&quot;/static/459ac78d88ceeca142303b1fdbe2217a/2bef9/densepose.png&quot; srcSet=&quot;/static/459ac78d88ceeca142303b1fdbe2217a/6f3f2/densepose.png 256w,/static/459ac78d88ceeca142303b1fdbe2217a/01e7c/densepose.png 512w,/static/459ac78d88ceeca142303b1fdbe2217a/2bef9/densepose.png 1024w,/static/459ac78d88ceeca142303b1fdbe2217a/71c1d/densepose.png 1536w,/static/459ac78d88ceeca142303b1fdbe2217a/a878e/densepose.png 2048w,/static/459ac78d88ceeca142303b1fdbe2217a/30a3b/densepose.png 3254w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;결국, 입력데이터와 출력 Patch(각 segmentation)을 잘 설계함으로써 2D 구조의 CNN으로 3D를 예측하는 방법을 고안해 낸 셈이다.&lt;/p&gt;&lt;h3 id=&quot;retinaface&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#retinaface&quot; aria-label=&quot;retinaface permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;RetinaFace&lt;/h3&gt;&lt;p&gt;Multiscale의 FPN을 이용해 다양한 task의 branch를 모두 넣어 다양한 task를 한번에 풀수 있도록 만든 모델이다(&lt;em&gt;FPN + Multi-task branches&lt;/em&gt;). 또는 Multitask 학습법이라고도 부른다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;classification, bounding box, 5 point regression, mesh regression&lt;/p&gt;&lt;p&gt;이 방식들은 조금씩 차이가 있지만 공통적으로 결국 얼굴에 대한 학습이므로, 이 모든 task들을 학습했을 때 좀 더 적은 데이터로도 backbone network 자체가 강인하게 학습되는 장점이 있다. &lt;/p&gt;&lt;p&gt;이러한 경향에서 볼 수 있는 현재 Computer Vision의 큰 흐름은, FPN에 Target-task에 해당하는 head branch만 만들어주어 다양한 응용 모델을 만드는 것이다.&lt;/p&gt;&lt;h1 id=&quot;detecting-objects-as-keypoints&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#detecting-objects-as-keypoints&quot; aria-label=&quot;detecting objects as keypoints permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Detecting objects as keypoints&lt;/h1&gt;&lt;p&gt;Bounding Box가 아닌 Key point로 object detection을 수행하는 방법에 대해 알아보자.&lt;/p&gt;&lt;h2 id=&quot;cornernet&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#cornernet&quot; aria-label=&quot;cornernet permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;CornerNet&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;CornerNet&lt;/strong&gt;은 Bounding Box가 (좌상단, 우하단) 단 두 개의 점좌표만 있으면 unique하게 결정될 수 있다는 점에 착안하여 고안되었다.&lt;/p&gt;&lt;/li&gt;&lt;/ul&gt;&lt;ol&gt;&lt;li&gt;&lt;p&gt;Backbone network(ConvNet)에서 나온 featuremap에 Heatmap을 통하여 각 bounding box의 corner pair를 추출한다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;각 포인터가 가진 정보를 표현하는 Embedding layer를 통과시킨다. 이 때, 같은 코너에서 나온 embedding point는 서로 같은 object를 표현하는 정보임을 학습시킨다.&lt;/p&gt;&lt;p&gt;성능보다는 좀 더 속도에 치중한 모델이다.&lt;/p&gt;&lt;h2 id=&quot;centernet&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#centernet&quot; aria-label=&quot;centernet permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;CenterNet&lt;/h2&gt;&lt;h3 id=&quot;1기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#1%EA%B8%B0&quot; aria-label=&quot;1기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;1기&lt;/h3&gt;&lt;p&gt;CornerNet의 성능을 조금 더 보완하기 위한 방법으로, (좌상단,우하단) 뿐만 아니라 (중앙) 좌표값도 사용한다.&lt;/p&gt;&lt;h3 id=&quot;2기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#2%EA%B8%B0&quot; aria-label=&quot;2기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;2기&lt;/h3&gt;&lt;p&gt;좀 더 정확하게 bounding box를 표현하기 위한 방법으로, (폭, 높이, 중앙)좌표값을 사용한다.&lt;/p&gt;&lt;hr/&gt;&lt;h3 id=&quot;reference&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#reference&quot; aria-label=&quot;reference permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Reference&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://medium.datadriveninvestor.com/yolact-real-time-instance-segmentation-5cbe6fc2ba36&quot;&gt;YOLACT - Real Time Instance Segmentation&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://analyticsindiamag.com/introduction-to-yolactedge-for-real-time-object-segmentation-on-edge-device/&quot;&gt;Introduction To YolactEdge For Real-time Object Segmentation On Edge Device&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://towardsdatascience.com/panoptic-segmentation-with-upsnet-12ecd871b2a3&quot;&gt;Panoptic Segmentation with UPSNet&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://www.youtube.com/watch?v=AW6ZMy8TWdI&amp;amp;ab_channel=ComputerVisionFoundationVideos&quot;&gt;Video Panoptic Segmentation&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://curt-park.github.io/2018-07-03/stacked-hourglass-networks-for-human-pose-estimation/&quot;&gt;[분석] Stacked Hourglass Networks for Human Pose Estimation&lt;/a&gt;&lt;/p&gt;&lt;/li&gt;&lt;/ol&gt;&lt;/li&gt;&lt;/ul&gt;</content:encoded></item><item><title><![CDATA[Computer Vision 07 - 조건부 생성 모델(Conditional generative model)]]></title><description><![CDATA[Conditional generative model by 오태현 교수님, BoostCamp AI Tech 7주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/43_conditional_generative_model/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/43_conditional_generative_model/</guid><pubDate>Thu, 11 Mar 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;conditional-generative-model&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#conditional-generative-model&quot; aria-label=&quot;conditional generative model permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Conditional generative model&lt;/h1&gt;&lt;p&gt;기존의 Generative 모델은 랜덤하게 샘플링만 가능했다. 그러나 단순히 생성에서 끝나는게 아니라, 원하는 모양대로 조건(condition)을 넣어 이에 맞는 모델을 생성할 수 있다면 훨씬 더 많은 응용가능성들이 있을 것이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;오디오 음질 향상, 기계번역, 제목을 이용해 뉴스 기사 작성 등&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Conditional GAN은 rough하게는 기존의 GAN에서 입력 input으로 Condition을 더 추가해주는 것으로 구현할 수 있다. 이미지 변환 기법으로도 다양한 사용 사례가 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Style transfer, Super resolution(화질 향상), Colorization ...&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;예시--super-resolution&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%98%88%EC%8B%9C--super-resolution&quot; aria-label=&quot;예시  super resolution permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;예시 : Super resolution&lt;/h2&gt;&lt;p&gt;입력으로 저해상도 영상이 주어지고, 출력으로 고해상도 영상이 나온다.&lt;/p&gt;&lt;p&gt;화질 향상 task는 사실 Conditional GAN으로만 해결할 수 있는 것은 아니고, 이전까지는 단순한 회귀 모델로 해결하려는 시도가 있었다. 저해상도 이미지를 넣어 generator가 고해상도 이미지를 출력하면, 실제 정답 (고해상도) 이미지와 MAE(L1) or MSE(L2) loss를 최소화하도록 학습하는 방식이었다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:34.375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;mae-mse-gan.png&quot; title=&quot;mae-mse-gan.png&quot; src=&quot;/static/e575065d90c49a5bc787489b2717d23e/2bef9/mae-mse-gan.png&quot; srcSet=&quot;/static/e575065d90c49a5bc787489b2717d23e/6f3f2/mae-mse-gan.png 256w,/static/e575065d90c49a5bc787489b2717d23e/01e7c/mae-mse-gan.png 512w,/static/e575065d90c49a5bc787489b2717d23e/2bef9/mae-mse-gan.png 1024w,/static/e575065d90c49a5bc787489b2717d23e/71c1d/mae-mse-gan.png 1536w,/static/e575065d90c49a5bc787489b2717d23e/a878e/mae-mse-gan.png 2048w,/static/e575065d90c49a5bc787489b2717d23e/30f63/mae-mse-gan.png 2816w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;일반적인 GAN의 generator &amp;#x27;정말 진짜같은가&amp;#x27;가 중요하기 때문에, 데이터를 넣었을 때 최대한 그 데이터에 가까운 형태를 뽑아내려고 한다. 즉, 실제 데이터셋에 너무 치우쳐서, 특정 형태의 정답 레이블과 거의 동일한 형태의 결과물을 만들어낸다. 이와 달리, &lt;strong&gt;&lt;div&gt;MAE/MSE based regression model은 학습에 사용된 모든 데이터들에 대해 비슷한 수준의 distance를 유지하는 방식으로 학습하므로, 어떤 데이터에도 치우치지 않고 중간영역을 선택한다.&lt;/div&gt;&lt;/strong&gt; 따라서 결국은 어정쩡한 average 형태의 이미지, 일종의 blurry한 이미지가 나오게 되는 것이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;예를 들어, 검은색/흰색 이미지를 가지고 학습했다고 하면, MAE/MSE 회귀 모델은 두 색상과의 distance를 모두 최소화해야하므로 어정쩡한 회색(average) 출력이 나와버린다. 그러나 GAN은 검은색이면 검은색, 흰색이면 흰색, 한 레이블에 치우쳐진 결과가 나올 것이다. 그렇게 해야 주어졌던 정답 레이블과 동일한 이미지를 만들 수 있기 때문이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Super resolution task에 이와 같은 MAE/MSE based regression model을 적용한 것이 &lt;strong&gt;SRResNet&lt;/strong&gt;이고, GAN 방식을 적용한 것이 &lt;code&gt;SRGAN&lt;/code&gt;이다.&lt;/p&gt;&lt;h1 id=&quot;image-translation-gans&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#image-translation-gans&quot; aria-label=&quot;image translation gans permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Image translation GANs&lt;/h1&gt;&lt;h2 id=&quot;pix2pix&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#pix2pix&quot; aria-label=&quot;pix2pix permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Pix2Pix&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Pix2Pix&lt;/code&gt;&lt;/strong&gt;는 Conditional Generative Model 중 하나로, 어떤 이미지를 다른 도메인의 상응하는 이미지로 생성하는 모델을 의미한다. 이미지 변환은 input과 output의 출력해상도가 유지되는 task인데, 이런 task를 CNN 기반으로 처음 해결한 방법이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;스타일 변환, segment to image, coloring 등...&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:29.6875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAAAsElEQVQY032P0Q6DIAxF/f+fnFkMiMuEVqCABdbpk8nmeYA+9HIPg7W8LGwMz4YBq4MO2Liy0Fqtlft/hhA2Im+M1np6zRpmtU7PVWtAfIxvpeAufF6t9VrPJm6AvZReeXMcsJSSiSjGmPN3SCnt+34JX8i5EfVE5FJyPsYgeO+VUuM46gNElIeGnz4iIoTQ0Tfx4QMplHIxOfxYzuHmS2LnLAOA1FprZRAF55zkz4UPqkpeBUCD/eYAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;pix2pix&quot; title=&quot;pix2pix&quot; src=&quot;/static/88978510f54006268d2f496a442b071f/2bef9/pix2pix-loss.png&quot; srcSet=&quot;/static/88978510f54006268d2f496a442b071f/6f3f2/pix2pix-loss.png 256w,/static/88978510f54006268d2f496a442b071f/01e7c/pix2pix-loss.png 512w,/static/88978510f54006268d2f496a442b071f/2bef9/pix2pix-loss.png 1024w,/static/88978510f54006268d2f496a442b071f/71c1d/pix2pix-loss.png 1536w,/static/88978510f54006268d2f496a442b071f/a878e/pix2pix-loss.png 2048w,/static/88978510f54006268d2f496a442b071f/ce034/pix2pix-loss.png 3314w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;code&gt;GAN Loss&lt;/code&gt;와 &lt;code&gt;L1(MAE) Loss&lt;/code&gt;를 모두 쓴다. 이유는 다음과 같다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;GAN Loss는 입력된 두개의 pair(input image와 정답 레이블)의 오차를 비교하는 것이 아니라, 단순히 real/fake 여부만 판단하기 때문에 y(정답 레이블)와 비슷하게 학습할 수 없다. 이러한 문제를 보완하기 위해 L1 Loss를 곁들여 쓴다. L1 Loss는 정답과의 차이를 최소화하고, GAN Loss는 실제같음을 추구한다.&lt;/li&gt;&lt;li&gt;GAN의 학습은 다른 모델에 비해 불안정하고 어려운 면이 있다. blurry한 이미지를 만들게 되거나, 학습이 잘 되지 않을 경우를 대비하여 L1 Loss도 같이 사용한다.&lt;/li&gt;&lt;/ol&gt;&lt;h2 id=&quot;cyclegan&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#cyclegan&quot; aria-label=&quot;cyclegan permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;CycleGAN&lt;/h2&gt;&lt;p&gt;Pix2Pix에서는 supervised learning이 들어가기 때문에, &amp;quot;pairwise data&amp;quot;가 필요했다. 그렇지만 pairwise dataset을 확보하는 것은 어렵다. 그래서 짝이 없는 데이터들도 활용할 수 있는 방법으로 &lt;strong&gt;&lt;code&gt;CycleGAN&lt;/code&gt;&lt;/strong&gt;이 제안되었다.&lt;/p&gt;&lt;p&gt;CycleGAN은 1:1 대응관계가 존재하지않는 데이터셋만으로도 traslation이 가능하도록 만든 모델이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;모네의 그림 ↔ 사진, 말 ↔ 얼룩말&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:17.96875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAECAIAAAABPYjBAAAACXBIWXMAABYlAAAWJQFJUiTwAAAAjUlEQVQI142O2w6DIBBE+f9v5KGGcFEU2CVqECidxvreCRmGsHNAKKW01ikl7/08z/BlWeDWWmYmGswjxu9CrnVA70ei944zEa3relOAcM7FGI0xRGyM3bYQQsQMMm7vCiTu7TxPlHPOrTXk4zgwDlDvbd/3UkrOPE0vKSV+6h/9yv8ID9Zar+sCHUT4BxkL5NfO4pmkAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;cyclegan-loss&quot; title=&quot;cyclegan-loss&quot; src=&quot;/static/8bc888878c743505b2a6f8d31c57b25c/2bef9/cyclegan-loss.png&quot; srcSet=&quot;/static/8bc888878c743505b2a6f8d31c57b25c/6f3f2/cyclegan-loss.png 256w,/static/8bc888878c743505b2a6f8d31c57b25c/01e7c/cyclegan-loss.png 512w,/static/8bc888878c743505b2a6f8d31c57b25c/2bef9/cyclegan-loss.png 1024w,/static/8bc888878c743505b2a6f8d31c57b25c/71c1d/cyclegan-loss.png 1536w,/static/8bc888878c743505b2a6f8d31c57b25c/a878e/cyclegan-loss.png 2048w,/static/8bc888878c743505b2a6f8d31c57b25c/351ee/cyclegan-loss.png 3350w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;Loss는 양방향의 GAN Loss와 &lt;strong&gt;&lt;em&gt;cycle-consistency loss&lt;/em&gt;&lt;/strong&gt;를 모두 사용한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;GAN Loss : 양방향으로 변환이 가능해야 하므로, X→Y, Y→X로 가는 방향 모두에 대해 동시에 학습을 수행한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mi&gt;Y&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;F&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L(D_X)+L(D_Y)+L(G)+L(F)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;Y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;F&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;F&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;G,F&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;F&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : Generator&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mi&gt;Y&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;D_X,D_Y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;Y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : Discriminator&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;Cycle-consistency Loss : 이미지를 translation하고, 그 결과물을 다시 원래 방향으로 translation했을 때 원본과 동일한 형태를 가지는가?&lt;/p&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;만약 GAN Loss만 사용한다면 어떻게 될까? 어떤 input이 들어가든지 현실과 비슷한 이미지를 생성만 해낸다면 Real로 판단할 것이므로, Input 이미지는 어떤 의미도 없어진다. 마치 local minimum에 빠지게 되는 것과 비슷하다.&lt;/p&gt;&lt;p&gt;따라서 단순히 실사와 비슷한 이미지를 생성하는것 뿐만 아니라, &amp;#x27;input 이미지라는 조건을 고려해야한다&amp;#x27;라는 제약을 두기 위해 Cycle-consistency Loss를 사용한다(preserve contents). 이 방식은 애초에 supervise가 필요하지 않다. 즉, self-supervised learning이다. 정답 레이블 자체가 자기자신이기 때문이다. Input 이미지를 변환시켜 돌아왔을 때 자기 스스로와 비교한다.&lt;/p&gt;&lt;h2 id=&quot;perceptual-loss&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#perceptual-loss&quot; aria-label=&quot;perceptual loss permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Perceptual loss&lt;/h2&gt;&lt;p&gt;GAN은 학습이 어렵다. 학습의 결과를 평가하기도 어렵고, 가장 쉬운 방식으로 자꾸 빠져버리는, 최빈값 함몰 문제(Mode Collapse)가 있기 때문이다. Alternating traing도 필요하다. 한번의 학습마다 Generator, Disciriminator를 각각 하나씩 만들어주어야 한다. 그럼에도 불구하고 GAN을 사용하는 이유는 퀄리티때문이다. L1 Loss나 L2 Loss를 사용하는 모델에 비해 훨씬 더 고퀄리티의 아웃풋을 낸다.&lt;/p&gt;&lt;p&gt;GAN을 사용하지 않고도 그 정도의 퀄리티를 내는 모델은 없을까? Perceptual loss는 그런 고퀄리티 아웃풋을 만들기 위해 나오게 된 loss이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;GAN Loss&lt;ul&gt;&lt;li&gt;학습시키기 어렵고, 코딩이 복잡하다.(Alternating training)&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Pre-trained Network가 필요하지 않다.&lt;/strong&gt; 학습과정에서 알아서 균형을 맞추기 때문이다.&lt;/li&gt;&lt;li&gt;기학습된 네트워크가 필요하지 않기 때문에 다양한 곳에 적용시킬 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Perceptual Loss&lt;ul&gt;&lt;li&gt;&lt;strong&gt;학습과 코딩(구현)이 쉽다. 간단한 순전파&amp;amp;역전파로 학습한다.&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;다만, loss를 measure하기 위해서 pre-trained network가 필요하다. 이를 learned loss라고 한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Pre-trained classifier의 필터 response가 인간의 시각적 지각과 유사하다는 관찰 결과가 있었다. 기학습된 네트워크의 초기 layer filter값을 보면, edge를 캐치하는 부분이나 color를 캐치하는 부분이 있다. 인간의 시각도 이와 마찬가지로 시각의 초기 단계에서 이런 부분들을 캐치한다고 한다. 따라서, Perceptual loss는 이미지를 인간이 세상을 바라보는 방식, perceptual space로 변환해서 바라볼 수 있지 않을까?라는 관점에서 시작되었다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;perceptual space는 인간이 시각 정보를 중요한 것과 중요하지 않은 것으로 구별하는 것처럼, 중요한 정보와 그렇지 않은 것이 나누어진 공간이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이런 Perceptual loss를 사용하면, GAN과 같이 복잡한 방식을 사용하지 않고도 loss 하나만 추가를 해서 style transfer가 가능한 Generator를 학습할 수 있다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:31.640625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA/ElEQVQY02WQXU+DMBSG+f8/xoQo8WLhbkRdows6vjsHoR3QjTaAARo+aukSNPqkOe3F+/Y952hCiHmexS8459eCsLLsur7nfDm9fC1wVVa9druGYWiaZhzHaZqGcUwsizpOXVdlnhUI13VTFoQyVlVVgdAao0l1lmVXBV9C+rbrPj8Oue8f7fDVenrfb5MDQLYV7ED4/JZ6oVDdLmaZBgCAEMpfpZ8qHM/1gsg0wL2+edw+AFPfm3c7w3jRN4l7M6tkOcARwjPGKEWEEIQQY9T28yCpMM7DCMZpjNNTBE9nnH3VDSvpT9viH23b2m7sR6mcYV0nuVBCLn+U33BdTTGhuBmIAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;perceptual-loss&quot; title=&quot;perceptual-loss&quot; src=&quot;/static/d082393e73a84a9d75ded822b22d8a26/2bef9/perceptual-loss.png&quot; srcSet=&quot;/static/d082393e73a84a9d75ded822b22d8a26/6f3f2/perceptual-loss.png 256w,/static/d082393e73a84a9d75ded822b22d8a26/01e7c/perceptual-loss.png 512w,/static/d082393e73a84a9d75ded822b22d8a26/2bef9/perceptual-loss.png 1024w,/static/d082393e73a84a9d75ded822b22d8a26/71c1d/perceptual-loss.png 1536w,/static/d082393e73a84a9d75ded822b22d8a26/a878e/perceptual-loss.png 2048w,/static/d082393e73a84a9d75ded822b22d8a26/f51c0/perceptual-loss.png 2994w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Image Transform Net : 주어진 input 데이터에 의존하여 (이를 토대로) 출력물을 생성한다.&lt;/li&gt;&lt;li&gt;Loss Network : 이미지 분류 문제로 &lt;strong&gt;기학습된 네트워크&lt;/strong&gt;를 사용하여, Image Transform Net의 출력물로부터 feature를 뽑아낸다. 이후 &lt;code&gt;Style Target/Content Target&lt;/code&gt;을 이용하여 출력물과의 loss를 토대로 각자 학습하여 새로운 스타일로 변환시켜준다.&lt;ul&gt;&lt;li&gt;&lt;strong&gt;Image Transform Net이 학습 중일때에는 Fix 되어있다(파라미터 update를 하지 않는다).&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;feature-reconstruction-loss&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#feature-reconstruction-loss&quot; aria-label=&quot;feature reconstruction loss permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Feature reconstruction loss&lt;/h3&gt;&lt;p&gt;Image Transform Net의 출력물(transformed image) &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\hat{y}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.19444em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 추가적으로 변환하는 과정에서, &lt;strong&gt;Content target&lt;/strong&gt;과 비교하여 input 이미지 x에 있던 content를 그대로 유지하고 있는지 확인시켜주는 loss를 &lt;code&gt;Feature reconstruction loss&lt;/code&gt;라고 한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:33.203125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAABGklEQVQY032Qy0oDQRBF528VBiQKQfARBPFBwI3gH0iyUPyHEBeuJIm7gIIOJkxPcB49PV093V1lJU7AIHjpRXOpU/dSQZ5T+kVK0W8hkV8/BECl2LHWOuesx8p6j2xQ0DnE9i4OBw0WfdB7hEXloHZgrHZk53M7nYKUaZYVWZZr81YYWbslfHmG56f4Mmng3i21duzNtSsl70aOMGWpZ591Vf10AmNEqXkFIgYnR3iwj0+PXJCX2fueDbdMuO2FWM4WhYzjWCSJEMIY471fsBLBJgAEV13sXuB41CQ/3NFxh/p9Ar0KwvUVVqJNBe09bIU4GPKRMFX6NapmC+285hyt4S+wAY+eaTKmUvLf83RdawOglGZxsf/hb3L+hh0DkuwlAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;feature-reconstruction-loss&quot; title=&quot;feature-reconstruction-loss&quot; src=&quot;/static/53da9ff2aa2f0fc66b9bb62ee01c2adf/2bef9/feature-reconstruction-loss.png&quot; srcSet=&quot;/static/53da9ff2aa2f0fc66b9bb62ee01c2adf/6f3f2/feature-reconstruction-loss.png 256w,/static/53da9ff2aa2f0fc66b9bb62ee01c2adf/01e7c/feature-reconstruction-loss.png 512w,/static/53da9ff2aa2f0fc66b9bb62ee01c2adf/2bef9/feature-reconstruction-loss.png 1024w,/static/53da9ff2aa2f0fc66b9bb62ee01c2adf/71c1d/feature-reconstruction-loss.png 1536w,/static/53da9ff2aa2f0fc66b9bb62ee01c2adf/a878e/feature-reconstruction-loss.png 2048w,/static/53da9ff2aa2f0fc66b9bb62ee01c2adf/365b1/feature-reconstruction-loss.png 3184w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;Content target에서 온 feature maps, transformed image에서 온 feature maps 두 개를 비교한 loss로 backpropagation을 수행한다. VGG Net 중간에서 얻어지는 feature map들은 어느정도 semantic한 의미를 담고있는 정보들이기 때문에, Input 이미지와 의미론적으로 비슷한지 체크할 수 있다.&lt;/p&gt;&lt;h3 id=&quot;style-reconstruction-loss&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#style-reconstruction-loss&quot; aria-label=&quot;style reconstruction loss permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Style reconstruction loss&lt;/h3&gt;&lt;p&gt;이번에는 변환시키고 싶은 형태의 image를 넣었을 때, 그 이미지의 Style feature map을 뽑고, 이를 통해 &lt;strong&gt;Gram matrix&lt;/strong&gt;를 만든다. 현재 이미지의 Gram matrix와 원하는 스타일 이미지의 Gram matrix를 비교한 Loss를 &lt;code&gt;Style reconstruction loss&lt;/code&gt;라고 한다. Feature reconstruction loss는 각 Feature map의 직접적인 비교를 통해 얻어냈지만, Style reconstruction loss는 Gram matrices를 비교한다는 데에서 차이점이 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Gram matrix가 왜 style을 비교하는 데에 쓰일 수 있는지에 대해서는 &lt;a href=&quot;https://www.facebook.com/groups/TensorFlowKR/permalink/512281009112962/&quot;&gt;이 게시물&lt;/a&gt;을 참고하도록 하자.&lt;/li&gt;&lt;li&gt;간략하게 말해서, Gram matrix는 &lt;strong&gt;feature map에서 spatial information를 배제하고, 이미지 전체의 전반적인 특징인 statistic만 포함하도록 만들었다&lt;/strong&gt;고 생각하면 된다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이런 Style reconstruction loss를 아예 사용하지 않고, Feature reconstruction loss + L2 loss 등을 이용하여 스타일 변환없이 퀄리티 향상을 하는 경우도 있다. 이 경우 content는 동일하게 유지한 상태로, 스타일은 변하지 않는다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Super resolution 등&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;various-gan-applications&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#various-gan-applications&quot; aria-label=&quot;various gan applications permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Various GAN Applications&lt;/h1&gt;&lt;h2 id=&quot;deepfake&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#deepfake&quot; aria-label=&quot;deepfake permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Deepfake&lt;/h2&gt;&lt;p&gt;GAN이나 Generative Model의 응용사례 중 가장 성공적인 사례는 Deepfake이다. 이미지나 비디오에서 사람의 얼굴, 목소리를 생성할 수 있는 기술이다.&lt;/p&gt;&lt;p&gt;그러나 Deepfake의 발전에 따라 윤리적 위험성이 크게 증가하여, 이를 detection하거나 예방할 수 있는 방법에 대한 많은 연구도 진행되고 있다.&lt;/p&gt;&lt;h2 id=&quot;face-de-identification&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#face-de-identification&quot; aria-label=&quot;face de identification permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Face de-identification&lt;/h2&gt;&lt;p&gt;GAN의 좋은 사례로, 사람의 익명성/privacy를 보호하는 가짜 이미지를 생성할 수도 있다. 사람의 얼굴을 조금만 변형하여, 사람들은 육안으로 누구인지 구별할 수 있지만 컴퓨터는 구별하지 못하는 가짜 얼굴을 만들 수 있다.&lt;/p&gt;&lt;p&gt;이를 응용하여, 올바른 패스워드를 넣었을 때에만 올바른 얼굴(original image)을 보여주고, 그렇지 않으면 변형된 다른 이미지를 넘겨주는 사례도 있다.(Face anonymization with passcode). 개인정보를 보호하는데에 사용된다.&lt;/p&gt;&lt;h2 id=&quot;video-translation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#video-translation&quot; aria-label=&quot;video translation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Video translation&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;Pose transfer : A의 외형과 B의 포즈를 차용하여 B의 포즈를 취하고 있는 A의 영상을 만들어내기&lt;/li&gt;&lt;li&gt;Video-to-Video translation : semantic segmentation map을 주면, 이에 상응하는 실사 영상을 만들어주는 것 → 게임에 활용 가능!&lt;/li&gt;&lt;/ul&gt;</content:encoded></item><item><title><![CDATA[Computer Vision 04 - 객체 검출(Object detection)]]></title><description><![CDATA[Object Detection by 오태현 교수님, BoostCamp AI Tech 7주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/40_object_detection/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/40_object_detection/</guid><pubDate>Wed, 10 Mar 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;p&gt;Object detection, 객체 검출은 CV에서 가장 시장수요가 높은 task 중 하나이다. 특히 자율주행(Autonomous driving), OCR(Optical Character Recognition) 등의 핵심 기술이다.&lt;/p&gt;&lt;p&gt;기존의 Semantic Segmentation과 [&lt;strong&gt;&lt;code&gt;Instance segmentation&lt;/code&gt;&lt;/strong&gt;, &lt;strong&gt;&lt;code&gt;Panoptic segementation&lt;/code&gt;&lt;/strong&gt;]의 차이점은, 전자는 class를 구별하지만 후자는 개체를 구별한다는 데에 있다(Instance를 구별하는가?). 예를 들어, 전자는 이미지에 나온 모든 사람들을 &amp;#x27;사람&amp;#x27;이라는 하나의 class로 묶어서 동일하게 취급하는데에 비하여, 후자는 Alex, Chris 등으로 각각 다른 사람으로 구별한다. 이중 Panoptic segemetation은 Instance segementation을 포함하고 있는 좀 더 큰 기술이다.&lt;/p&gt;&lt;p&gt;객체 검출(Object detection)은 이미지 분류(Image Classification)와 Box Localization의 조합이다. 일반 영상인식보다 좀 더 고차원의 task로, 먼저 Box의 좌표(좌상단, 우하단)를 잡아 객체를 찾아낸 뒤, 해당 객체의 카테고리를 분류한다.&lt;/p&gt;&lt;h1 id=&quot;기존-기법--수작업하기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B8%B0%EC%A1%B4-%EA%B8%B0%EB%B2%95--%EC%88%98%EC%9E%91%EC%97%85%ED%95%98%EA%B8%B0&quot; aria-label=&quot;기존 기법  수작업하기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;기존 기법 : 수작업하기&lt;/h1&gt;&lt;h2 id=&quot;gradient-based-detector&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gradient-based-detector&quot; aria-label=&quot;gradient based detector permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Gradient-based detector&lt;/h2&gt;&lt;p&gt;&amp;quot;객체를 검출하기 위해, 경계선을 따면 좋지 않을까?&amp;quot;라는 생각으로 경계선 검출에 집중했다. 영역 내의 수평선/수직선 등의 분포를 모델링하고, linear classifier의 weight를 학습시켜 visualization했다. 예를 들어, &amp;#x27;어깨부분에 사선이 많고, 팔다리 부분에 수직선이 많은 사진은 사람으로 판별하라&amp;#x27;라는 등의 로직이다.&lt;/p&gt;&lt;p&gt;feature를 뽑아내는 부분은 사람이 직접 디자인하므로 매우 정교하고, 학습가능한 linear 모델의 weight부는 상대적으로 비중이 적은 셈이었다.&lt;/p&gt;&lt;h2 id=&quot;selective-search&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#selective-search&quot; aria-label=&quot;selective search permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Selective Search&lt;/h2&gt;&lt;p&gt;&lt;img src=&quot;https://media.geeksforgeeks.org/wp-content/uploads/home/step3-660x304.PNG&quot; alt=&quot;selective-search&quot;/&gt;&lt;/p&gt;&lt;p&gt;최근까지도 많이 사용한 기술로, 사람이나 특정 물체 뿐만 아니라, 다양한 물체 후보군의 영역 후보군을 지정해주는 방식이다. 즉, 수많은 Bounding Box(BB)를 제안해준다.&lt;/p&gt;&lt;p&gt;제안받은 여러 BB중 가장 객체의 영역을 tight하게 잘 잡아내는 Box를 찾도록 학습시킨다. 딥러닝을 이용한 초기 물체 탐지에도 이 방식이 많이 사용되었다.&lt;/p&gt;&lt;p&gt;실제로 딥러닝 기반의 Object detection 기법들은 2014년 이후 나오기 시작했으며, 크게 두 갈래로 분류할 수 있다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;&lt;code&gt;Two-stage detector&lt;/code&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;Single-stage detector&lt;/code&gt;&lt;/li&gt;&lt;/ol&gt;&lt;h1 id=&quot;two-stage-detector&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#two-stage-detector&quot; aria-label=&quot;two stage detector permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Two-stage detector&lt;/h1&gt;&lt;h2 id=&quot;r-cnn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#r-cnn&quot; aria-label=&quot;r cnn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;R-CNN&lt;/h2&gt;&lt;p&gt;2012년 AlexNet의 Image Classification 네트워크가 압도적인 성능을 보여주자, 바로 Object detection에 응용되었다. 기존 방법 대비 압도적으로 높은 성능을 보여주며 객체 검출영역에 데뷔한 방식이다.&lt;/p&gt;&lt;p&gt;&lt;em&gt;Regions with CNN features&lt;/em&gt;의 약자로, CNN에 Region proposal 방식을 도입하였다.&lt;/p&gt;&lt;h3 id=&quot;작동-과정&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9E%91%EB%8F%99-%EA%B3%BC%EC%A0%95&quot; aria-label=&quot;작동 과정 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;작동 과정&lt;/h3&gt;&lt;ol&gt;&lt;li&gt;이미지를 입력한다.&lt;/li&gt;&lt;li&gt;약 2000개(2k) 이하로 region proposal(BB 후보군)을 추출한다.&lt;/li&gt;&lt;li&gt;각 region proposal을 CNN input에 적절한 크기로 warping(이미지 사이즈 일정하게 조정)을 해준다&lt;/li&gt;&lt;li&gt;target task가 아닌 다른 task에 대해 학습된 Pre-trained CNN에 넣는다.&lt;/li&gt;&lt;li&gt;CNN의 FC layer에서 추출된 feature를 기반으로 SVM의 linear classifier만을 이용해서 클래스를 학습한다.(fine-tuning)&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;그러나, 이 방식은 2천개나 되는 box에 대해 일일이 다 수행을 해주므로 &lt;strong&gt;굉장히 느리다&lt;/strong&gt;는 문제가 있었다. 또, region proposal은 selective search같은 수작업 알고리즘 기반이라, &lt;strong&gt;학습이 불가능하다&lt;/strong&gt;는 한계가 있었다.&lt;/p&gt;&lt;h2 id=&quot;fast-r-cnn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#fast-r-cnn&quot; aria-label=&quot;fast r cnn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Fast R-CNN&lt;/h2&gt;&lt;p&gt;따라서 R-CNN의 단점을 보완하기 위해 영상 전체에 대한 feature를 한번에 추출하고, 미리 검출된 이 feature들을 재활용하여 객체를 검출하는 &lt;strong&gt;&lt;code&gt;Fast R-CNN&lt;/code&gt;&lt;/strong&gt;이 등장했다.&lt;/p&gt;&lt;p&gt;&lt;img src=&quot;%5Bhttps://www.learnopencv.com/wp-content/uploads/2019/06/frcnn.png&quot; alt=&quot;fast-r-cnn&quot;/&gt;&lt;/p&gt;&lt;h3 id=&quot;작동방식&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9E%91%EB%8F%99%EB%B0%A9%EC%8B%9D&quot; aria-label=&quot;작동방식 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;작동방식&lt;/h3&gt;&lt;ol&gt;&lt;li&gt;원본 이미지에서 Convolution layer까지 feature map을 미리 검출한다.&lt;ul&gt;&lt;li&gt;이 시점의 feature map은 conv를 거쳤으므로 tensor형태(C,H,W)가 된다.&lt;/li&gt;&lt;li&gt;Fully convolutional Network는 입력 사이즈를 따로 warping하지 않아도 feature map을 추출할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;한번 뽑아놓은 feature를 여러번 재활용하기 위해 region proposal이 제시한 물체들의 후보군, 즉&lt;strong&gt;&lt;code&gt;RoI(Region of Interest)&lt;/code&gt;&lt;/strong&gt;에 해당하는 feature만을 추출한다.&lt;/li&gt;&lt;li&gt;RoI feature를 고정된 사이즈(fixed size)로 resampling한다. 이 과정을 &lt;em&gt;RoI pooling layer&lt;/em&gt;라고 한다.&lt;/li&gt;&lt;li&gt;이후, resampling된 RoI feature를 다음의 두 과정에 동시&lt;ol&gt;&lt;li&gt;이미지 분류 : softmax에 통과시킨다.&lt;/li&gt;&lt;li&gt;정확한 BB 크기 찾기 : bbox regressor에 통과시킨다.&lt;/li&gt;&lt;/ol&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;이렇게 feature만 재활용했는데도 기존의 R-CNN에 비해 18배나 빠른 수준이었다. 그러나 region proposal은 아직도 selective search를 했으므로, 학습이 불가능한 부분이 수작업 알고리즘이 존재하여, 아무리 데이터가 많더라도 여전히 한계가 존재했다. &lt;/p&gt;&lt;h2 id=&quot;faster-r-cnn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#faster-r-cnn&quot; aria-label=&quot;faster r cnn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Faster R-CNN&lt;/h2&gt;&lt;p&gt;앞선 두 R-CNN의 최대 단점인 Selective search를 제거하고, 그 부분을 NN으로 대체하였다. 따라서 object detection 분야에서 &lt;strong&gt;최초의 end-to-end 모델&lt;/strong&gt;(모든 네트워크의 컴포넌트 NN기반)이 되었다.&lt;/p&gt;&lt;h3 id=&quot;iou와-anchor-boxes&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#iou%EC%99%80-anchor-boxes&quot; aria-label=&quot;iou와 anchor boxes permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;IOU와 Anchor boxes&lt;/h3&gt;&lt;p&gt;&lt;code&gt;IOU(Intersection over Union)&lt;/code&gt;는 두 영역의 overlap을 측정하는 기준을 제공하는 metric이다. &lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mtext&gt;IoU =&lt;/mtext&gt;&lt;mfrac&gt;&lt;mtext&gt;Area of Overlap(교집합)&lt;/mtext&gt;&lt;mtext&gt;Area of Union(합집합)&lt;/mtext&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\text {IoU =}\frac{\text{Area of Overlap(교집합)}}{\text{Area of Union(합집합)}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.363em;vertical-align:-0.936em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;IoU =&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;Area of Union(&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;합집합&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;Area of Overlap(&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;교집합&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;![anchor-box](&lt;a href=&quot;https://www.researchgate.net/profile/Max-Ferguson/publication/327392506/figure/fig8/AS:666613162450944@1535944371721/Anchor-Boxes-at-a-certain-position-in-the-feature-map.png%5D(https://www.researchgate.net/profile/Max-Ferguson/publication/327392506/figure/fig8/AS:666613162450944@1535944371721/Anchor-Boxes-at-a-certain-position-in-the-feature-map.png)&quot;&gt;https://www.researchgate.net/profile/Max-Ferguson/publication/327392506/figure/fig8/AS:666613162450944@1535944371721/Anchor-Boxes-at-a-certain-position-in-the-feature-map.png](https://www.researchgate.net/profile/Max-Ferguson/publication/327392506/figure/fig8/AS:666613162450944@1535944371721/Anchor-Boxes-at-a-certain-position-in-the-feature-map.png)&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;code&gt;Anchor boxes&lt;/code&gt;는 각 위치에서 발생할 것 같은 box를 미리 rough하게 정의해놓은(pre-defined) 후보군이다. 위치마다, 크기마다 미리 정해놓은 box들을 가져다 붙이며 쓴다. 그 중 IoU가 0.7을 넘어가는 anchor box의 경우 positive sample, 즉 정답으로 간주하고, 0.3을 넘어가지 않는 box는 negative sample로 두고 penalty를 주도록 학습했다. box의 개수와 종류는 하이퍼파라미터이지만, Faster R-CNN에서는 총 9개를 설정했다. 즉, BB에 어떻게 loss를 적용시킬 지 결정하는 기준이다.&lt;/p&gt;&lt;h3 id=&quot;작동방식-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9E%91%EB%8F%99%EB%B0%A9%EC%8B%9D-1&quot; aria-label=&quot;작동방식 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;작동방식&lt;/h3&gt;&lt;p&gt;시간이 오래 걸리고, 일종의 3rd-party 알고리즘이었던 Selective search 방식을 &lt;strong&gt;&lt;code&gt;Region Proposal Network(RPN)&lt;/code&gt;&lt;/strong&gt;으로 대체하였다.&lt;/p&gt;&lt;p&gt;![faster r-cnn](&lt;a href=&quot;https://www.researchgate.net/profile/Giang-Son-Tran/publication/324549019/figure/fig1/AS:649929152266241@1531966593689/Faster-R-CNN-Architecture-9.png%5D(https://www.researchgate.net/profile/Giang-Son-Tran/publication/324549019/figure/fig1/AS:649929152266241@1531966593689/Faster-R-CNN-Architecture-9.png)&quot;&gt;https://www.researchgate.net/profile/Giang-Son-Tran/publication/324549019/figure/fig1/AS:649929152266241@1531966593689/Faster-R-CNN-Architecture-9.png](https://www.researchgate.net/profile/Giang-Son-Tran/publication/324549019/figure/fig1/AS:649929152266241@1531966593689/Faster-R-CNN-Architecture-9.png)&lt;/a&gt;&lt;/p&gt;&lt;ol&gt;&lt;li&gt;Fast R-CNN과 마찬가지로, 영상 하나에서 나오는 feature를 미리 뽑아둔다.&lt;/li&gt;&lt;li&gt;RPN에서 region proposal을 여러가지 제공한다.&lt;/li&gt;&lt;li&gt;region proposal을 바탕으로 RoI Pooling을 수행하고, 그 결과를 바탕으로 classify한다.&lt;/li&gt;&lt;/ol&gt;&lt;h3 id=&quot;rpn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#rpn&quot; aria-label=&quot;rpn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;RPN&lt;/h3&gt;&lt;p&gt;&lt;img src=&quot;https://miro.medium.com/max/375/1*JDQw0RwmnIKeRABw3ZDI7Q.png&quot; alt=&quot;rpn&quot;/&gt;&lt;/p&gt;&lt;p&gt;conv feature map을 256D로 뽑아내면, 다음의 두 score를 만들어낸다&lt;/p&gt;&lt;ol&gt;&lt;li&gt;cls layer - object 여부를 판단하는 2k개의 classification scores&lt;ul&gt;&lt;li&gt;각각의 anchor box에 대해 object인지/아닌지를 계산하므로 2k개&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;reg layer - k개 anchor box의 정교한 위치를 회귀하는 4k개의 bounding box regression output&lt;ul&gt;&lt;li&gt;BB 하나를 정하기 위해서 한 꼭짓점의 (x,y)좌표, 너비 w, 높이 h 총 4개의 변수가 필요하므로 4k개&lt;/li&gt;&lt;li&gt;anchor box가 아주 촘촘하면 이렇게 하지 않아도 되겠지만, 그러면 계산속도가 아주 느려질 것이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ol&gt;&lt;h3 id=&quot;non-maximum-suppressionnms&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#non-maximum-suppressionnms&quot; aria-label=&quot;non maximum suppressionnms permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Non-Maximum Suppression(NMS)&lt;/h3&gt;&lt;p&gt;아무리 objectiveness score를 판단한다고 해도, 객체로 판단되는 BB가 너무 많을 수 있다. &lt;code&gt;NMS&lt;/code&gt;는 이 중 하나를 정확히 특정하기 위해서 그럴듯한 BB를 모두 가져다놓고 다른 박스와의 IOU를 측정해 너무 많이 겹치는(ex-IoU≥50%) 것들을 모두 제거하는 방식이다. 이 기술은 딥러닝 이전의 방식에서도 대부분 사용되었지만, 딥러닝에서도 BB를 필터링하기 위해 사용되는 standard한 알고리즘이다.&lt;/p&gt;&lt;h1 id=&quot;single-stage-detector&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#single-stage-detector&quot; aria-label=&quot;single stage detector permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Single-stage detector&lt;/h1&gt;&lt;p&gt;two-stage detecor와는 다르게, 정확도를 조금 포기하더라도 속도를 올리는 데에 초점을 맞춘 &lt;strong&gt;real-time detector&lt;/strong&gt;이다. &lt;strong&gt;&lt;div&gt;region proposal을 기반으로 한 RoI pooling을 사용하지 않고 곧바로 box regression과 classification만 사용하므로, 구조가 비교적 간단하고 속도가 훨씬 빠르다.&lt;/div&gt;&lt;/strong&gt;&lt;/p&gt;&lt;h2 id=&quot;you-only-look-onceyolo&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#you-only-look-onceyolo&quot; aria-label=&quot;you only look onceyolo permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;You Only Look Once(YOLO)&lt;/h2&gt;&lt;h3 id=&quot;작동방식-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9E%91%EB%8F%99%EB%B0%A9%EC%8B%9D-2&quot; aria-label=&quot;작동방식 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;작동방식&lt;/h3&gt;&lt;p&gt;&lt;img src=&quot;https://miro.medium.com/max/700/1*9nikM2b0u-m67SJpQXftKA.png&quot; alt=&quot;YOLO&quot;/&gt;&lt;/p&gt;&lt;ol&gt;&lt;li&gt;이미지를 그리드로 나눈다.&lt;/li&gt;&lt;li&gt;각 grid에 대해서 score를 예측한다.&lt;ol&gt;&lt;li&gt;B개의 Bounding box마다 박스 정보 4개(x,y,w,h)와 confident score(objectiveness score) 1개를 예측한다.&lt;/li&gt;&lt;li&gt;class score(class probability)도 따로 예측한다.&lt;/li&gt;&lt;/ol&gt;&lt;/li&gt;&lt;li&gt;최종 결과는 NMS를 통해 확정한다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;&lt;img src=&quot;https://lilianweng.github.io/lil-log/assets/images/yolo-network-architecture.png&quot; alt=&quot;YOLO2&quot;/&gt;&lt;/p&gt;&lt;p&gt;S : 마지막 conv layer의 해상도&lt;/p&gt;&lt;p&gt;B : anchor box의 개수&lt;/p&gt;&lt;p&gt;C : class의 개수&lt;/p&gt;&lt;h2 id=&quot;single-shot-multibox-detectorssd&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#single-shot-multibox-detectorssd&quot; aria-label=&quot;single shot multibox detectorssd permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Single Shot MultiBox Detector(SSD)&lt;/h2&gt;&lt;p&gt;YOLO의 prediction은 마지막 layer에서 단 한번만 수행하기 때문에, Faster R-CNN에 비해 localization 정확도가 조금 떨어지는 아쉬움이 있었다. 이를 보완하기 위해 &lt;code&gt;SSD&lt;/code&gt;가 나왔다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:50.390625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAKCAIAAAA7N+mxAAAACXBIWXMAABYlAAAWJQFJUiTwAAABZklEQVQoz21Ri26EIBD0/z+xTe6uZ1QEAXmDoNvRSy9t0gkSk92ZWWY7IjourOsaQsy57juVreHgP6WCw5iYmVDKxXhYG40x64UOtJxzKUVrPY1sHhbnq5R+5mbs1743EwtcJMY853Wa6NmbmXPnPFy7RUoogQ+VlFJrW86plJxiiN57m7SMSm1K76s5XKDhNj5ut6/+ieZOLIu1FuR9351ztbVwsoKxLl2Agl998GUeV8bj/eP5vD/6YQClg/u2wS231mKMLwJmMZC0Dh+ErHNSSbkoLdU8sZEx1M+x34G9AAkIwSiEoPW6yAVZnDylFqHmYbbaHHS2neRaa4gRgcE/pQzb1xT0g9rIFcqeqqIiqbp3hTpwQEa30prPnDHGOYct/YKaxNdjwH7oLzrkBCv403/AQ3Drx/3z8wMxHG1vKWO3dPV3WJoQAtkgrFdgSBc3ROsFNAVr8WbUIkbivAwDlbK39g2FhD55V1J06gAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;ssd&quot; title=&quot;ssd&quot; src=&quot;/static/e1e6dfe6dcce59f0c5a5de21c111b711/2bef9/ssd.png&quot; srcSet=&quot;/static/e1e6dfe6dcce59f0c5a5de21c111b711/6f3f2/ssd.png 256w,/static/e1e6dfe6dcce59f0c5a5de21c111b711/01e7c/ssd.png 512w,/static/e1e6dfe6dcce59f0c5a5de21c111b711/2bef9/ssd.png 1024w,/static/e1e6dfe6dcce59f0c5a5de21c111b711/71c1d/ssd.png 1536w,/static/e1e6dfe6dcce59f0c5a5de21c111b711/a878e/ssd.png 2048w,/static/e1e6dfe6dcce59f0c5a5de21c111b711/ea66f/ssd.png 2828w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;SSD는 Multiscale object를 더 잘 처리하기 위하여 중간 feature map을 여러 해상도에 맞추어 출력할 수 있도록 만들었다. Conv 연산을 거치면서 각 scale마다 feature map이 출력되는데, 이를 모두 활용하여 각 scale마다 다양한 크기의 object에 맞추어 대응할 수 있도록 만들어졌다.&lt;/p&gt;&lt;p&gt;YOLO보다 빠른 속도와 더 좋은 성능, 심지어 Faster R-CNN까지 제쳐버리며 압도적인 퍼포먼스를 달성했다(다만 Input size가 달랐다)&lt;/p&gt;&lt;h1 id=&quot;two-stage-detector-vs-single-stage-detector&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#two-stage-detector-vs-single-stage-detector&quot; aria-label=&quot;two stage detector vs single stage detector permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Two-stage detector vs Single-stage detector&lt;/h1&gt;&lt;h2 id=&quot;focal-loss&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#focal-loss&quot; aria-label=&quot;focal loss permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Focal loss&lt;/h2&gt;&lt;h3 id=&quot;class-imbalance-problem&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#class-imbalance-problem&quot; aria-label=&quot;class imbalance problem permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Class imbalance problem&lt;/h3&gt;&lt;p&gt;Single-stage detector들은 RoI Pooling을 포기했는데, 이 때문에 모든 영역에서의 loss가 계산되고 그에 따라 일정 gradient가 발생하게 된다. 일반적인 영상의 경우 검출하고자 하는 객체의 면적은 얼마 되지 않고, 나머지 대부분은 background이다. 이말은 즉, 정보가 많은 positive sample은 굉장히 적고, 유용한 정보가 없는 negative sample은 굉장히 많이 나온다는 말이다.&lt;/p&gt;&lt;p&gt;이런 &lt;code&gt;class imbalance&lt;/code&gt; 문제를 해결하기 위해 &lt;code&gt;focal loss&lt;/code&gt;라는 것이 도입되었는데, cross-entropy의 확장이라고 볼 수 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mtext&gt;CE&lt;/mtext&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mtext&gt;FL&lt;/mtext&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;γ&lt;/mi&gt;&lt;/msup&gt;&lt;/mstyle&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
\text{CE}(p_t) &amp;amp;= -\log(p_t) \\
\text{FL}(p_t) &amp;amp;= -\textcolor{red}{(1-p_t)^\gamma}\log(p_t)
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.0000000000000004em;vertical-align:-1.2500000000000002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7500000000000002em&quot;&gt;&lt;span style=&quot;top:-3.91em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;CE&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.41em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;FL&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2500000000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7500000000000002em&quot;&gt;&lt;span style=&quot;top:-3.91em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.41em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot; style=&quot;color:red&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.714392em&quot;&gt;&lt;span style=&quot;top:-3.1130000000000004em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05556em;color:red&quot;&gt;γ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2500000000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;Cross Entropy에서 확률 term이 추가되었는데, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;γ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\gamma&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05556em&quot;&gt;γ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 잘 맞추었는지 여부에 따라 결정된다. 잘 맞추었을 경우 loss를 더 적게 주고, 못 맞추었을 경우 loss를 더 크게 주어 penalty를 강화하는 것이다. 즉, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;γ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\gamma&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05556em&quot;&gt;γ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 클수록 훨씬 더 sharp하게 변화하게 된다.&lt;/p&gt;&lt;h2 id=&quot;retinanet&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#retinanet&quot; aria-label=&quot;retinanet permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;RetinaNet&lt;/h2&gt;&lt;p&gt;&lt;img src=&quot;https://miro.medium.com/max/2268/1*0-GVAp6WCzPMR6puuaYQTQ.png&quot; alt=&quot;retinanet&quot;/&gt;&lt;/p&gt;&lt;p&gt;Low level의 특징 layer들과 High level의 특징을 둘 다 잘 활용하면서도 각 scale별로 물체를 잘 찾기위한 multi scale구조로 만들어졌다. &lt;/p&gt;&lt;p&gt;중간중간 feature map들을 넘겨주고, 그 값들을 모두 더해준다(concat이 아니라 add로 fusion한다). 이후, class head와 box head가 따로 구별이 되어서 classification과 box regression을 각 위치마다 dense하게 수행된다.&lt;/p&gt;&lt;p&gt;SSD와 속도는 비슷하지만, 성능은 더 나은 모습을 보였다.&lt;/p&gt;&lt;h1 id=&quot;detection-with-transformer&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#detection-with-transformer&quot; aria-label=&quot;detection with transformer permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Detection with Transformer&lt;/h1&gt;&lt;p&gt;NLP에서 큰 성공을 거둔 Transformer가 컴퓨터 비젼에 접목되어서 새로운 연구 결과들이 나오고 있다.&lt;/p&gt;&lt;h2 id=&quot;detection-transformerdetr&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#detection-transformerdetr&quot; aria-label=&quot;detection transformerdetr permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Detection Transformer(DETR)&lt;/h2&gt;&lt;p&gt;페이스북에서 2020년 발표한 모델로, Objecte detection에 Transformer를 적용시킨 사례이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:37.109375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;detr&quot; title=&quot;detr&quot; src=&quot;/static/24447230826ee789975c3ae21ca52fa6/2bef9/detr.png&quot; srcSet=&quot;/static/24447230826ee789975c3ae21ca52fa6/6f3f2/detr.png 256w,/static/24447230826ee789975c3ae21ca52fa6/01e7c/detr.png 512w,/static/24447230826ee789975c3ae21ca52fa6/2bef9/detr.png 1024w,/static/24447230826ee789975c3ae21ca52fa6/71c1d/detr.png 1536w,/static/24447230826ee789975c3ae21ca52fa6/a878e/detr.png 2048w,/static/24447230826ee789975c3ae21ca52fa6/89896/detr.png 2966w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ol&gt;&lt;li&gt;CNN에서 뽑아낸 feature와, input image를 각 위치의 multi dimension으로 표현한 positional encoding을 쌍으로 하여 입력 토큰을 생성한다.(position-해당 position의 feature)&lt;/li&gt;&lt;li&gt;Transformer encoder를 통과시키고, 정리된 feature들을 decoder에 넣어준다. 이 때, object queries도 같이 넣어 해당 위치에 어떤 feature가 존재하는지를 decoder의 출력값으로 받는다. &lt;ul&gt;&lt;li&gt;object queries : 학습된 위치 인코딩(Learned positional encodings for querying)&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;받아낸 출력값을 토대로 detection 파트(prediction heads)에서 box를 어떻게 그려야하는지 출력해주게 된다. &lt;/li&gt;&lt;/ol&gt;&lt;p&gt;이 때, 하이퍼파라미터로 N을 넣어주게 되는데, 이것은 한 이미지 내에서 검출될 수 있는 최대 object 갯수를 의미한다.&lt;/p&gt;&lt;hr/&gt;&lt;h3 id=&quot;reference&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#reference&quot; aria-label=&quot;reference permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Reference&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://www.geeksforgeeks.org/selective-search-for-object-detection-r-cnn/&quot;&gt;Selective Search for Object Detection | R-CNN - GeeksforGeeks&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://www.geeksforgeeks.org/fast-r-cnn-ml/&quot;&gt;Fast R-CNN | ML - GeeksforGeeks&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://www.researchgate.net/figure/Anchor-Boxes-at-a-certain-position-in-the-feature-map_fig8_327392506&quot;&gt;Anchor Boxes at a certain position in the feature map - fig 8&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://www.researchgate.net/figure/Faster-R-CNN-Architecture-9_fig1_324549019&quot;&gt;Faster R-CNN Architecture 9&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://medium.com/egen/region-proposal-network-rpn-backbone-of-faster-r-cnn-4a744a38d7f9&quot;&gt;Region Proposal Network (RPN) - Backbone of Faster R-CNN&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://towardsdatascience.com/yolov1-you-only-look-once-object-detection-e1f3ffec8a89&quot;&gt;Review: YOLOv1 - You Only Look Once (Object Detection)&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://towardsdatascience.com/review-retinanet-focal-loss-object-detection-38fba6afabe4&quot;&gt;Review: RetinaNet - Focal Loss (Object Detection)&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[Computer Vision 05 - CNN 시각화(Visualization)]]></title><description><![CDATA[Semantic segmentation by 오태현 교수님, BoostCamp AI Tech 7주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/41_cnn_visualization/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/41_cnn_visualization/</guid><pubDate>Wed, 10 Mar 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;visualizing-cnn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#visualizing-cnn&quot; aria-label=&quot;visualizing cnn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Visualizing CNN&lt;/h1&gt;&lt;p&gt;CNN 학습 시에 어떤 과정으로 정확한 prediction을 학습하게 되는지는 블랙박스로 남아있다. 또, 성능이 잘 나오지 않는다면 어떤 이유로 학습이 잘 되지 않았는지도 알기 어렵다. 이러한 점에서 CNN의 내부 시각화의 필요성이 생긴다. 시각화 툴은 일종의 디버깅 툴처럼 사용될 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;em&gt;ZFNet&lt;/em&gt;이 시각화 툴을 사용했던 사례&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;vanilla-example--filter-visualization&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#vanilla-example--filter-visualization&quot; aria-label=&quot;vanilla example  filter visualization permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Vanilla example : filter visualization&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:47.265625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;filter-visualization&quot; title=&quot;filter-visualization&quot; src=&quot;/static/4761076305f51040ed891783e29743d6/2bef9/filter-visualization.png&quot; srcSet=&quot;/static/4761076305f51040ed891783e29743d6/6f3f2/filter-visualization.png 256w,/static/4761076305f51040ed891783e29743d6/01e7c/filter-visualization.png 512w,/static/4761076305f51040ed891783e29743d6/2bef9/filter-visualization.png 1024w,/static/4761076305f51040ed891783e29743d6/71c1d/filter-visualization.png 1536w,/static/4761076305f51040ed891783e29743d6/a878e/filter-visualization.png 2048w,/static/4761076305f51040ed891783e29743d6/01dfe/filter-visualization.png 2910w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;가장 간단한 시각화 방법으로는, 필터 자체를 시각화하거나 convolution을 취한 activation map도 시각화할 수 있다. 다만, 초기 layer에는 catch하는 영역이 명확하게 보이지만, 중-후기 layer로 갈수록 앞쪽 layer의 필터들과 합성되어서 더 추상적인 정보를 detect하므로 사실상 사람이 해석가능한 정보가 별로 없어 시각화의 의미가 없다.&lt;/p&gt;&lt;h3 id=&quot;신경망-시각화의-유형type&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%8B%A0%EA%B2%BD%EB%A7%9D-%EC%8B%9C%EA%B0%81%ED%99%94%EC%9D%98-%EC%9C%A0%ED%98%95type&quot; aria-label=&quot;신경망 시각화의 유형type permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;신경망 시각화의 유형(type)&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:33.984375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;types-of-visualization&quot; title=&quot;types-of-visualization&quot; src=&quot;/static/1292914289391c820793d7614dc9e126/2bef9/types-of-visualization.png&quot; srcSet=&quot;/static/1292914289391c820793d7614dc9e126/6f3f2/types-of-visualization.png 256w,/static/1292914289391c820793d7614dc9e126/01e7c/types-of-visualization.png 512w,/static/1292914289391c820793d7614dc9e126/2bef9/types-of-visualization.png 1024w,/static/1292914289391c820793d7614dc9e126/71c1d/types-of-visualization.png 1536w,/static/1292914289391c820793d7614dc9e126/a878e/types-of-visualization.png 2048w,/static/1292914289391c820793d7614dc9e126/bc81a/types-of-visualization.png 2862w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;크게 &lt;code&gt;model behavior&lt;/code&gt; 분석과 &lt;code&gt;model decision&lt;/code&gt; 분석으로 나눌 수 있으며, 모델에 집중하느냐 데이터에 집중하느냐의 차이이다.&lt;/p&gt;&lt;h1 id=&quot;analysis-of-model-behaviors&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#analysis-of-model-behaviors&quot; aria-label=&quot;analysis of model behaviors permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Analysis of model behaviors&lt;/h1&gt;&lt;h2 id=&quot;embedding-feature-analysis&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#embedding-feature-analysis&quot; aria-label=&quot;embedding feature analysis permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Embedding feature analysis&lt;/h2&gt;&lt;p&gt;High-level layer에서 얻는 High-level feature들을 분석하는 방법이다.&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;nearest-neighborsnn-in-a-feature-space&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#nearest-neighborsnn-in-a-feature-space&quot; aria-label=&quot;nearest neighborsnn in a feature space permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Nearest neighbors(NN) in a feature space&lt;/h3&gt;&lt;p&gt;이전에 배웠던 Nearest neighbor 방법을 활용하여, DB에 수많은 사진들을 넣어놓고 특정 Query image를 날려 해당 image와 비슷한 feature space의 images를 찾아내는지 확인한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:46.484375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;nn-visualization&quot; title=&quot;nn-visualization&quot; src=&quot;/static/d79d05716f9e1e6763673932f99035bd/2bef9/nn-visualization.png&quot; srcSet=&quot;/static/d79d05716f9e1e6763673932f99035bd/6f3f2/nn-visualization.png 256w,/static/d79d05716f9e1e6763673932f99035bd/01e7c/nn-visualization.png 512w,/static/d79d05716f9e1e6763673932f99035bd/2bef9/nn-visualization.png 1024w,/static/d79d05716f9e1e6763673932f99035bd/71c1d/nn-visualization.png 1536w,/static/d79d05716f9e1e6763673932f99035bd/a878e/nn-visualization.png 2048w,/static/d79d05716f9e1e6763673932f99035bd/4a70a/nn-visualization.png 2390w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;이 이미지를 이용해 다음과 같이 판단할 수 있다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;눈으로 직접 보고, 의미론적으로(semantically) 비슷한 이미지들이 클러스터링되어있는지 확인한다.&lt;/li&gt;&lt;li&gt;간단하게, query image와 모델이 찾아낸 neighbor image 간을 픽셀별로 distance 계산하여 비슷한 이미지인지 확인한다.&lt;ul&gt;&lt;li&gt;이 경우는 위치가 다르거나 포즈가 다른 이미지들을 올바르게 판단할 수 없는 경우가 많다.&lt;/li&gt;&lt;li&gt;만약 포즈와 위치가 다른 이미지들도 명확하게 neighbor로 분류한다면, 이는 모델이 위치변화에 강인하게, 컨셉을 제대로 학습했다는 의미이므로 모델의 강건성(robustness)을 입증하는 결과가 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;그러나, 이 방식은 예제를 보고 판단해야 하므로 전체적인 조감도를 보기는 어렵다는 단점이 있다.&lt;/p&gt;&lt;h3 id=&quot;dimensionality-reduction&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#dimensionality-reduction&quot; aria-label=&quot;dimensionality reduction permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Dimensionality reduction&lt;/h3&gt;&lt;p&gt;너무 고차원의 feature space를 상상하거나 눈으로 확인하여 판단하기 어렵기 때문에, 차원을 축소하여 분포를 확인하는 방법이다. 대표적인 방법으로 &lt;code&gt;t-distributed stochastic neighbor embedding(t-SNE)&lt;/code&gt; 가 있다.&lt;/p&gt;&lt;p&gt;&lt;img src=&quot;https://kr.mathworks.com/help/examples/stats/win64/VisualizeHighDimensionalDataUsingTSNEExample_01.png&quot; alt=&quot;https://kr.mathworks.com/help/examples/stats/win64/VisualizeHighDimensionalDataUsingTSNEExample_01.png&quot;/&gt;&lt;/p&gt;&lt;p&gt;모여있는 형태를 보고, 튀는 데이터들과 모델이 어느 클래스들을 유사하게 생각하는지 등에 대한 정보를 알 수 있다.&lt;/p&gt;&lt;h2 id=&quot;activation-investigation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#activation-investigation&quot; aria-label=&quot;activation investigation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Activation investigation&lt;/h2&gt;&lt;p&gt;Middle~High level를 해석하는 해석방법이다.&lt;/p&gt;&lt;h3 id=&quot;layer-activation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#layer-activation&quot; aria-label=&quot;layer activation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Layer activation&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:78.90625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;layer-activation&quot; title=&quot;layer-activation&quot; src=&quot;/static/9a8498a6e14b49bd4838c8d84039b49f/2bef9/layer-activation.png&quot; srcSet=&quot;/static/9a8498a6e14b49bd4838c8d84039b49f/6f3f2/layer-activation.png 256w,/static/9a8498a6e14b49bd4838c8d84039b49f/01e7c/layer-activation.png 512w,/static/9a8498a6e14b49bd4838c8d84039b49f/2bef9/layer-activation.png 1024w,/static/9a8498a6e14b49bd4838c8d84039b49f/71c1d/layer-activation.png 1536w,/static/9a8498a6e14b49bd4838c8d84039b49f/0d292/layer-activation.png 1620w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위의 사진을 보면, conv5 layer에서 hidden unit 중 하나를 뽑아서 thresholding하여 masking했더니 어떤 노드는 얼굴만 찾아내고, 어떤 노드는 계단만 찾아내는 것을 확인할 수 있다. 이처럼 학습 과정에서 특정 노드들이 어떤 역할을 하고있는지를 시각화하여 볼 수 있다. 이런 점을 보았을 때, CNN은 층을 쌓으면서 중간의 노드들이 여러 객체(얼굴, 계단 등)의 detection 역할을 담당하고, 결과적으로 이 detection들을 합쳐서 객체를 검출해내는 역할을 한다고 해석할 여지가 생긴다.&lt;/p&gt;&lt;h3 id=&quot;maximally-activating-patches&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#maximally-activating-patches&quot; aria-label=&quot;maximally activating patches permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Maximally activating patches&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:55.46875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;Maximally-activating-patches&quot; title=&quot;Maximally-activating-patches&quot; src=&quot;/static/bd1bf909fd9450dacdc822fe7b5dddc2/2bef9/maximally-activating-patches.png&quot; srcSet=&quot;/static/bd1bf909fd9450dacdc822fe7b5dddc2/6f3f2/maximally-activating-patches.png 256w,/static/bd1bf909fd9450dacdc822fe7b5dddc2/01e7c/maximally-activating-patches.png 512w,/static/bd1bf909fd9450dacdc822fe7b5dddc2/2bef9/maximally-activating-patches.png 1024w,/static/bd1bf909fd9450dacdc822fe7b5dddc2/71c1d/maximally-activating-patches.png 1536w,/static/bd1bf909fd9450dacdc822fe7b5dddc2/a878e/maximally-activating-patches.png 2048w,/static/bd1bf909fd9450dacdc822fe7b5dddc2/1192d/maximally-activating-patches.png 2662w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위 이미지는 CNN의 각 층에서 가장 높은 값을 가지는 hidden unit 주위의 영역을 뜯어낸 patch들의 사진인데, 이를 보면 해당 unit이 어떤 역할을 하고 있는지 확인할 수 있다. 어떤 유닛은 강아지의 코를, 어떤 유닛들은 색깔이 들어간 글자를 찾는 역할 등을 한다. 이 경우는 전반적인 큰 그림 보다는 국부적인 patch를 보므로 비교적 middle-level의 해석에 좋을 것이다.&lt;/p&gt;&lt;p&gt;과정은 다음과 같다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;특정 layer에서 channel 하나를 고른다.&lt;/li&gt;&lt;li&gt;예제 데이터를 backbone network에 넣어서 각 layer의 activation map을 모두 뽑고, 골랐던 채널의 activation map을 저장한다.&lt;/li&gt;&lt;li&gt;최대 activation value 주위의 이미지 패치를 잘라낸다(crop).&lt;ul&gt;&lt;li&gt;해당 value가 커버하는 receptive field를 찾아서 그것을 주위 영역으로 판단한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ol&gt;&lt;h3 id=&quot;class-visualization&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#class-visualization&quot; aria-label=&quot;class visualization permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Class visualization&lt;/h3&gt;&lt;p&gt;예제 데이터를 사용하지 않고, 네트워크가 기억(내재)하고 있는 이미지를 시각화하여 판단하는 방법도 있다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:102.34375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;class-visualization&quot; title=&quot;class-visualization&quot; src=&quot;/static/e1c7981840b71de8af790fb82e073fdf/2bef9/class-visualization.png&quot; srcSet=&quot;/static/e1c7981840b71de8af790fb82e073fdf/6f3f2/class-visualization.png 256w,/static/e1c7981840b71de8af790fb82e073fdf/01e7c/class-visualization.png 512w,/static/e1c7981840b71de8af790fb82e073fdf/2bef9/class-visualization.png 1024w,/static/e1c7981840b71de8af790fb82e073fdf/985a9/class-visualization.png 1512w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위의 이미지에서는 새(bird) 클래스와 개 클래스에 대한 네트워크의 예상치를 확인한 것이다. 이 때, 우하단을 보면 개를 제외하고도 아이의 형태가 나온것을 확인할 수 있는데, 이를 통해 클래스 분류에 단순히 해당 객체만 파악하는 것이 아니라, 주변 객체와의 연관성도 파악한다고 해석할 여지가 있다. 또한, 학습데이터에서 개가 대부분 아이와 등장했다는 것을 의미하므로, 학습 데이터의 편향성도 의심해볼 수 있다.&lt;/p&gt;&lt;p&gt;이 방법은 최적화를 통해 구현해야하는데, 형태는 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;I&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;/msup&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;arg max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mi&gt;I&lt;/mi&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;I&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;I&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;I&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mi&gt;I&lt;/mi&gt;&lt;msubsup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;I^* = \underset{I}{\argmax}f(I) - Reg(I)\\
Reg(I) = \lambda\Vert I\Vert^2_2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.738696em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;I&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.738696em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mbin mtight&quot;&gt;∗&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.688771em;vertical-align:-0.938771em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.161229em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07847em&quot;&gt;I&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938771em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;I&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.00773em&quot;&gt;R&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;I&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.00773em&quot;&gt;R&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;I&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1141079999999999em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;λ&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;I&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;I&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;I&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;I&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 영상 입력&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;I&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f(I)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;I&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 입력이 CNN 모델을 거쳐 나온 class score(ex- 개 클래스의 score)&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;I&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Reg(I)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.00773em&quot;&gt;R&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;I&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 임의의 argmax 결과를 사용하다 보면, 추상적인 값이 공교롭게도 가장 큰 값일 때가 있다. 이를 우리가 이해할 수 있는 형태로 바꾸어주기 위하여 간단한 loss인 정규화 term을 추가한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이전에는 Loss를 최소화하는 gradient descent를 사용했지만, 여기서는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;I&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f(I)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;I&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 최대화(argmax)되도록 해야하므로 Gradient ascent를 사용한다.&lt;/p&gt;&lt;p&gt;Gradient ascent의 과정은 다음과 같다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;임의의 영상(dummy image)을 CNN에 넣어 타겟 클래스의 prediction socre를 얻는다.&lt;/li&gt;&lt;li&gt;Backpropagation으로 입력 image의 gradient를 얻는다.&lt;ul&gt;&lt;li&gt;입력이 어떻게 변해야 target score가 높아지는지 찾는다&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Target score가 높아지는 방향으로 input image를 update해준다.(즉, gradient를 더해준다)&lt;/li&gt;&lt;li&gt;업데이트된 영상을 input image로 삼아 1-3을 반복한다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;이 때 최초의 dummy image는 여러 종류를 선택할 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;blank / monotone / noisy image 등..&lt;/li&gt;&lt;li&gt;이 때 최초에 넣어준 dummy image를 base로 입력 영상을 업데이트하므로, dummy image의 형태를 어느정도 따라가는 최종 image를 얻게된다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;model-decision-explanation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#model-decision-explanation&quot; aria-label=&quot;model decision explanation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Model decision explanation&lt;/h1&gt;&lt;p&gt;지금까지는 모델 자체의 행동을 분석했다면, 이번엔 모델이 특정 입력을 어떤 각도로 바라보고 있는지 살펴보자.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;saliency-test&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#saliency-test&quot; aria-label=&quot;saliency test permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Saliency test&lt;/h2&gt;&lt;p&gt;영상이 주어졌을 때, 영상이 제대로 판정되기 위한 각 영역의 중요도를 추출하는 방법이다.&lt;/p&gt;&lt;h3 id=&quot;occlusion-map&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#occlusion-map&quot; aria-label=&quot;occlusion map permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Occlusion map&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:42.1875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAICAIAAAB2/0i6AAAACXBIWXMAABYlAAAWJQFJUiTwAAABiUlEQVQY002QO0skQRSFWxQEf4SJICLspouJf0AwMTHQRMzNhWWjDd1EQTBRUDARNDAYRXHWZXRY3QFlHB/oKOg8urq66eqqflR3V11v9zbioepSp6jv3kMZAEBt6+q6VqvfPDZv0SqltNbwITxrla8QQAJEEEt0KAN3/Xzv+8L07vLs6f5qokBwTik1Cel2u8L3pQhCO447lejXEJsaTbaHk79fuHtZwO7T+friZG1jXjydYketUlSUCyOgw8v49YT/MFqDfWzMSKu9onVYwI5tVY93zn6Xms1GHvNT5synWW2V046hnJG0Mpwe9ETVcgHf/juYmxhZ+zlTLm3KWHPPI4SYptludxhjSmWT07tKsNUvrK9R8s0tDbwcHhUwvf+zsjB+vL3kNC/yyfhYhWGAylL8D5IE4uGSv1U5qbBOjRJSwDoJPdN8brx0257t4OdrSsFxPiWPIjAJuBwsD4gHlIFlQRgaUko/CJhQTIDHAWGLgm1j1VKmKlfG+z7Y2JJmlVpaCLx7B1uPqCHBLRmFAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;occlusion-map&quot; title=&quot;occlusion-map&quot; src=&quot;/static/51b018223a0e88fb60f12a25d6fe85bb/2bef9/occlusion-map.png&quot; srcSet=&quot;/static/51b018223a0e88fb60f12a25d6fe85bb/6f3f2/occlusion-map.png 256w,/static/51b018223a0e88fb60f12a25d6fe85bb/01e7c/occlusion-map.png 512w,/static/51b018223a0e88fb60f12a25d6fe85bb/2bef9/occlusion-map.png 1024w,/static/51b018223a0e88fb60f12a25d6fe85bb/71c1d/occlusion-map.png 1536w,/static/51b018223a0e88fb60f12a25d6fe85bb/a878e/occlusion-map.png 2048w,/static/51b018223a0e88fb60f12a25d6fe85bb/f7859/occlusion-map.png 3484w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;input image를 넣을 때, occlusion patch를 이용하여 가려준다. 이 때, occlusion patch를 넣은 이미지를 주었을 때 해당 이미지를 정답으로 판별할 확률(CNN score)을 계산하여, 패치가 어떤 위치를 가리고 있느냐에 따라 이 값이 바뀌는 정도를 기록해둔다.&lt;/p&gt;&lt;p&gt;위의 이미지로 보았을 때, score를 heatmap으로 표시하면, 어두운 영역이 물체의 검출에 민감한 요인이 되는 중요 영역임을 알 수 있다.&lt;/p&gt;&lt;h3 id=&quot;via-backpropagation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#via-backpropagation&quot; aria-label=&quot;via backpropagation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;via Backpropagation&lt;/h3&gt;&lt;p&gt;gradient ascent를 이용하여 클래스 이미지를 생성했던 방법과 비슷한데, &lt;strong&gt;이번엔 random image가 아니라 특정 image를 입력해 해당 image를 classification하는 데에 큰 영향을 끼친 부분을 heatmap으로 표시&lt;/strong&gt;한다.&lt;/p&gt;&lt;p&gt;과정은 다음과 같다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;타겟 입력 이미지를 CNN에 넣어 class score를 얻는다.&lt;/li&gt;&lt;li&gt;Backpropagation으로 입력 이미지의 gradient를 구한다.&lt;/li&gt;&lt;li&gt;얻어진 gradient에 절댓값을 취하거나, 제곱을 하여 절대적인 크기(magnitude)를 구한다.&lt;ul&gt;&lt;li&gt;어느 쪽으로 바뀌는지보다 해당 영역이 얼마나 큰 영향을 끼치는가가 중요하므로, 부호를 버리고 magnitude를 측정한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;해당 gradient magnitude map을 시각화한다. 필요에 따라 1-3을 반복하여 accumulate할수도 있다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;Class visualization의 gradient ascent와 다른 것은, 아무 이미지나 넣어 해당 클래스에 대한 모델의 예상치 추측을 하는것이 아니라, &lt;strong&gt;&lt;div&gt;특정 이미지에 대한 모델의 판단 요인을 찾는다는 것이다.&lt;/div&gt;&lt;/strong&gt; 즉, 현재 데이터가 어떻게 해석되는지를 보고싶은 것이므로 data-dependent하다.&lt;/p&gt;&lt;h2 id=&quot;backpropagation-based-saliency&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#backpropagation-based-saliency&quot; aria-label=&quot;backpropagation based saliency permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Backpropagation-based saliency&lt;/h2&gt;&lt;h3 id=&quot;deconvnet&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#deconvnet&quot; aria-label=&quot;deconvnet permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Deconvnet&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:81.640625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;deconvnet&quot; title=&quot;deconvnet&quot; src=&quot;/static/030d2fe01975a275d5248087f0d7bf3b/2bef9/deconvnet.png&quot; srcSet=&quot;/static/030d2fe01975a275d5248087f0d7bf3b/6f3f2/deconvnet.png 256w,/static/030d2fe01975a275d5248087f0d7bf3b/01e7c/deconvnet.png 512w,/static/030d2fe01975a275d5248087f0d7bf3b/2bef9/deconvnet.png 1024w,/static/030d2fe01975a275d5248087f0d7bf3b/71c1d/deconvnet.png 1536w,/static/030d2fe01975a275d5248087f0d7bf3b/d3d45/deconvnet.png 1614w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;일반적으로 CNN의 Forward에서는 Activation function으로 ReLU가 많이 사용된다. 따라서 음수가 0으로 마스킹되는데, 문제는 Backpropagation을 할 때에도 현재 값들을 기준으로 음수를 마스킹하지 않고 Forward 시에 0이하였던 unit들을 음수 마스킹해버린다는 것이다. &lt;strong&gt;&lt;em&gt;Deconvnet&lt;/em&gt;&lt;/strong&gt; 연산은 Backward 시에 Forward시점의 값 기준이 아니라 Backpropagation 시점의 값을 기준으로 음수를 마스킹한다. 이는 마치 역방향으로 ReLU를 재적용시킨 것과 같다. 이 경우 휴리스틱하게 좀 더 saliency를 잘 추출할 수 있었다고 한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mtext&gt;Forward Pass : &lt;/mtext&gt;&lt;msup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mtext&gt;Backward - backpropagation : &lt;/mtext&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;/msup&gt;&lt;mo&gt;&amp;gt;&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mtext&gt;Backward - deconvnet:&lt;/mtext&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo&gt;&amp;gt;&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
\text{Forward Pass : }h^{l+1} &amp;amp;= \max(0,h^l)\\
\text{Backward - backpropagation : }\frac{\partial L}{\partial h^l} &amp;amp;= [(h^l&amp;gt;0)]\frac{\partial L}{\partial h^{l+1}}\\
\text{Backward - deconvnet:}\frac{\partial L}{\partial h^l} &amp;amp;= [(h^{l+1}&amp;gt;0)]\frac{\partial L}{\partial h^{l+1}}
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:6.273988000000001em;vertical-align:-2.886994000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.386994em&quot;&gt;&lt;span style=&quot;top:-5.859325999999999em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;Forward Pass : &lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.827886em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;Backward - backpropagation : &lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7751079999999999em&quot;&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.470445999999999em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;Backward - deconvnet:&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7751079999999999em&quot;&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.886994000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.386994em&quot;&gt;&lt;span style=&quot;top:-5.859325999999999em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;max&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.827886em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;&amp;gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7751079999999999em&quot;&gt;&lt;span style=&quot;top:-2.989em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.470445999999999em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;&amp;gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7751079999999999em&quot;&gt;&lt;span style=&quot;top:-2.989em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.886994000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;h3 id=&quot;guided-backpropagation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#guided-backpropagation&quot; aria-label=&quot;guided backpropagation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Guided backpropagation&lt;/h3&gt;&lt;p&gt;여기서 더 한발 나아가서, Backward 시에 Forward 패턴도 마스킹하고, 현재 패턴 기준으로도 마스킹하는 방식을 &lt;strong&gt;&lt;em&gt;Guided backpropagation&lt;/em&gt;&lt;/strong&gt; 이라고한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;/msup&gt;&lt;mo&gt;&amp;gt;&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;&amp;amp;&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo&gt;&amp;gt;&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{\partial L}{\partial h^l} = [(h^l &amp;gt;0)\&amp;amp;(h^{l+1}&amp;gt;0)]\frac{\partial L}{\partial h^{l+1}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.05744em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7751079999999999em&quot;&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.149108em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;&amp;gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.149108em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&amp;amp;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;&amp;gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.05744em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7751079999999999em&quot;&gt;&lt;span style=&quot;top:-2.989em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;사실 이러한 방식은 수학적으로 크게 논리적인 방식은 아니나, 결과값으로부터 이 방식이 어떤 의미를 갖는지는 해석해볼 수 있다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:30.859375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;guided-backprop&quot; title=&quot;guided-backprop&quot; src=&quot;/static/2faae3a6a86cdf24b7cd2524f7e8e834/2bef9/guided-backprop.png&quot; srcSet=&quot;/static/2faae3a6a86cdf24b7cd2524f7e8e834/6f3f2/guided-backprop.png 256w,/static/2faae3a6a86cdf24b7cd2524f7e8e834/01e7c/guided-backprop.png 512w,/static/2faae3a6a86cdf24b7cd2524f7e8e834/2bef9/guided-backprop.png 1024w,/static/2faae3a6a86cdf24b7cd2524f7e8e834/71c1d/guided-backprop.png 1536w,/static/2faae3a6a86cdf24b7cd2524f7e8e834/a878e/guided-backprop.png 2048w,/static/2faae3a6a86cdf24b7cd2524f7e8e834/b0096/guided-backprop.png 2928w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;일반적인 backpropagation 방식으로 Forward 패턴의 양수(즉, 정답을 올바르게 예측하는데에 영향을 미쳤던 feature)도 참조하고, deconvnet 방식에서 Backward 패턴에서의 양수도 참조한다면, 결과적으로 두 과정에서 도움이 되는 feature들을 모아 보여주므로 조금 더 깨끗한 이미지가 나오게 되는 것이라고 해석 가능하다.&lt;/p&gt;&lt;h2 id=&quot;class-activation-mapping&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#class-activation-mapping&quot; aria-label=&quot;class activation mapping permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Class Activation Mapping&lt;/h2&gt;&lt;h3 id=&quot;class-activation-mapcam&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#class-activation-mapcam&quot; aria-label=&quot;class activation mapcam permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Class Activation Map(CAM)&lt;/h3&gt;&lt;p&gt;&lt;img src=&quot;https://i.imgur.com/PW2gSBgl.png&quot; alt=&quot;https://i.imgur.com/PW2gSBgl.png&quot;/&gt;&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;CAM&lt;/code&gt;&lt;/strong&gt; 아키텍쳐는 CNN의 일부를 개조하여 만들어진다. CNN의 conv파트를 최종적으로 통과하고 FC layer에 진입하기 전, 즉 마지막 conv feature map을 대상으로 global average pooling을 수행하여 Gap feature를 얻는다. 이후, 단 하나의 FC layer만 통과시켜 classification한다. 마지막으로 이미지 분류 task에 대해 학습을 다시 수행한다. 즉, 일종의 pretrained된 CNN 모델로부터 유도하는 아키텍쳐에 가깝다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/munder&gt;&lt;msubsup&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msubsup&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;msub&gt;&lt;mi&gt;F&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/munder&gt;&lt;msubsup&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msubsup&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;munder&gt;&lt;munder&gt;&lt;mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/munder&gt;&lt;msubsup&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mrow&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
S_c &amp;amp;= \sum_kw^c_k\textcolor{blue}{F_k}\\
&amp;amp;= \sum_kw^c_k\textcolor{blue}{\sum_{(x,y)}f_k(x,y)} = \sum_{(x,y)}\textcolor{red}{\underbrace{\sum_kw^c_kf_k(x,y)}_{CAM_c(x,y)}}
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:6.852236000000001em;vertical-align:-3.1761180000000007em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.6761180000000007em&quot;&gt;&lt;span style=&quot;top:-5.676118000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.05764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.024em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.1761180000000007em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.6761180000000007em&quot;&gt;&lt;span style=&quot;top:-5.676118000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000005em&quot;&gt;&lt;span style=&quot;top:-1.847887em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3021129999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7143919999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:blue&quot;&gt;F&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:blue&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.024em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000005em&quot;&gt;&lt;span style=&quot;top:-1.847887em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3021129999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7143919999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mopen mtight&quot; style=&quot;color:blue&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot; style=&quot;color:blue&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:blue&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot; style=&quot;color:blue&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;color:blue&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em;color:blue&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:blue&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:blue&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:blue&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:blue&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:blue&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000003em&quot;&gt;&lt;span style=&quot;top:-0.3748920000000006em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em;color:red&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em;color:red&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.16454285714285719em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.10903em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot; style=&quot;color:red&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:red&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500050000000005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000005em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-1.0998920000000005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;color:red;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;color:red;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;color:red;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;color:red;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500050000000005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000005em&quot;&gt;&lt;span style=&quot;top:-1.847887em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:red&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;color:red&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3021129999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;color:red;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em;color:red&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7143919999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:red&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em;color:red&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:red&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:red&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;color:red;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:red&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.950113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.850113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.1761180000000007em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;S_c&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.05764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 클래스 c에 대한 score값&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 마지막 conv layer의 채널 수.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;w^c_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.9474999999999999em;vertical-align:-0.2831079999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.664392em&quot;&gt;&lt;span style=&quot;top:-2.4168920000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2831079999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 마지막 FC layer에서 클래스 c에 해당하는 weight&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;F&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;F_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;F&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 채널별 conv feature map을 공간축(여러 채널)에 대하여 global average pooling한 것&lt;ul&gt;&lt;li&gt;이 때, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;F&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;F_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;F&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 결국 모든 픽셀 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(x,y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해서, conv feature map을 각 채널 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;마다 평균 취한 것이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sum_{(x,y)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.22471em;vertical-align:-0.47471em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.22528999999999993em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.47471em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : Global average pooling&lt;/li&gt;&lt;li&gt;모든 연산들이 선형 연산이므로 순서를 바꾸어줄 수 있다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;CAM_c(x,y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10903em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 결국 global average pooling을 적용하기 전이므로, 아직 공간에 대한 정보가 남아있다. 이것을 영상처럼 처리해서 visualization하면, 위 이미지의 하단 히트맵처럼 나오게 된다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;나오는 히트맵이 다른 방법에 비해 부드럽고 압도적으로 좋은 성능을 보여주기 때문에 자주 이용한다.&lt;/p&gt;&lt;p&gt;이미지 분류를 할 때, 위치에 대한 어떤 annotation 정보도 주지 않았는데도 위치까지 어느정도 찾아주기 때문에 bounding box를 만들어 object detection 등을 추가적으로 하는데에 사용되기도 한다.&lt;/p&gt;&lt;p&gt;이처럼 object detection과 같은 비교적 정교한 task를 좀 더 rough한 영상인식 task로 학습하여 처리하는 방식을 &lt;strong&gt;&lt;em&gt;weakly supervised learning&lt;/em&gt;&lt;/strong&gt; 이라고 부른다.&lt;/p&gt;&lt;p&gt;다만, CAM 적용을 위해서는 마지막 layer가 GAP과 FC layer로 이루어져야만 하며, 아키텍쳐를 바꾸고 나서 재학습을 해야한다는 단점이 있다. 이 경우 기학습된 모델에 비해 전체적 성능이 떨어지는 결과가 나올수도 있다. 최종 출력 이전에 Global average pooling과 FC layer 층이 이미 존재하므로 아키텍쳐를 수정하지 않고도 바로 CAM을 추출하기에 용이한 ResNet이나 GoogLeNet과 같은 사례들도 있다.&lt;/p&gt;&lt;h2 id=&quot;grad-cam&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#grad-cam&quot; aria-label=&quot;grad cam permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Grad-CAM&lt;/h2&gt;&lt;p&gt;CAM은 최종 층의 구조를 바꿔야 해서 모든 아키텍쳐에 적용할 수는 없다는 제약사항이 있었으므로, 구조를 변경하지 않고 기학습된 네트워크에서 CAM을 뽑을 수 있는 &lt;strong&gt;&lt;code&gt;Grad-CAM&lt;/code&gt;&lt;/strong&gt; 방식이 제안되었다.&lt;/p&gt;&lt;p&gt;&lt;img src=&quot;https://miro.medium.com/max/1186/0*D4FATkIeWp61o9zo.jpg&quot; alt=&quot;https://miro.medium.com/max/1186/0*D4FATkIeWp61o9zo.jpg&quot;/&gt;&lt;/p&gt;&lt;p&gt;기존 pretrained된 모델 아키텍쳐를 변경할 필요가 없기 때문에, 영상 인식 task에 한정될 필요가 없어졌다. 오로지Backbone이 CNN이기만 하면 사용할 수 있다. &lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/munder&gt;&lt;msub&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msup&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msup&gt;&lt;/mstyle&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sum_k\textcolor{red}{w^c}_kf_k(x,y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.3521180000000004em;vertical-align:-1.3021129999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000005em&quot;&gt;&lt;span style=&quot;top:-1.847887em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3021129999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em;color:red&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7143919999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;기존의 CAM 식에서 알아내야 하는 부분은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;w_k^c&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.9474999999999999em;vertical-align:-0.2831079999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.664392em&quot;&gt;&lt;span style=&quot;top:-2.4168920000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2831079999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, 즉 importance weight 뿐이므로, 이것을 알아내는 것이 핵심이다. &lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:31.640625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAABJklEQVQY03WQO07DQBCGfQBoOBVX4AgcgAIKJFwg0SJES4XgBAjRICFFgUBw3iRZJ8F5GJM4u/ZmduN4XziYggJ+TTH6Nd+vmbEYZ291Z+Ai8y1K6XQ2E0IkySqVKjeV1lKqOE7nOJnjFFjKOM98C6Hm3u7OiX3YdfsDzyu/Oi/P5ZHvFx6feu2OSRITEUHmnAMhxBsOx+NJGIaj0XgNu6hzfHRwd3tDgXnuoOY0qqgdttpuofjZQRm5CJf+RH0EBkBnGyy5jKnEZLUAY11dB5sbl7ZdoYDr1Val1mp6PSFkmiildBafigwzwIzW+RHrJi/r7PR9a/t8/8IBTGkU3z8USw0viiCYEiXlz9Q/sjDhzaAXRHT9GCG7qO+HmAEAY/oX9mfAF1o9RtFUDh+2AAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;grad-cam1&quot; title=&quot;grad-cam1&quot; src=&quot;/static/93c78cc8e1a831fad066e55bc26cbce2/2bef9/grad-cam1.png&quot; srcSet=&quot;/static/93c78cc8e1a831fad066e55bc26cbce2/6f3f2/grad-cam1.png 256w,/static/93c78cc8e1a831fad066e55bc26cbce2/01e7c/grad-cam1.png 512w,/static/93c78cc8e1a831fad066e55bc26cbce2/2bef9/grad-cam1.png 1024w,/static/93c78cc8e1a831fad066e55bc26cbce2/71c1d/grad-cam1.png 1536w,/static/93c78cc8e1a831fad066e55bc26cbce2/a878e/grad-cam1.png 2048w,/static/93c78cc8e1a831fad066e55bc26cbce2/b7535/grad-cam1.png 3556w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;Saliency를 Backprop으로 구했던 방법을 응용해서 수행한다.&lt;/p&gt;&lt;p&gt;여기서의 weight는 기존의 weight와 조금 다른 개념이기 때문에 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;α&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\alpha&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.0037em&quot;&gt;α&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 하자.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;기존의 Saliency test는 입력영상까지 backprop했지만, 여기에서는 원하는 activation map(즉, 특정 conv 층)까지만 backprop한다.&lt;/li&gt;&lt;li&gt;클래스 c에 대한 정답 레이블 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y^c&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.858832em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.664392em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로부터 Loss를 구한다.&lt;/li&gt;&lt;li&gt;이렇게 구한 weight가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;a_k^c&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.9474999999999999em;vertical-align:-0.2831079999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.664392em&quot;&gt;&lt;span style=&quot;top:-2.4168920000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2831079999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 된다.&lt;/li&gt;&lt;/ul&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;/mrow&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msubsup&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;U&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/munder&gt;&lt;msubsup&gt;&lt;mi&gt;α&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msubsup&gt;&lt;msup&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L^c_{Grad-CAM} = ReLU(\sum_k\alpha_k^cA^k)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.019723em;vertical-align:-0.305331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.714392em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.1130000000000004em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.305331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.3521180000000004em;vertical-align:-1.3021129999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.00773em&quot;&gt;R&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;U&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000005em&quot;&gt;&lt;span style=&quot;top:-1.847887em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3021129999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.0037em&quot;&gt;α&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7143919999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.0037em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;새로이 구한 weight &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;a_k^c&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.9474999999999999em;vertical-align:-0.2831079999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.664392em&quot;&gt;&lt;span style=&quot;top:-2.4168920000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2831079999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 activation map &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A^k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.849108em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.849108em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 선형결합하여 ReLU를 적용한다. 따라서, 양수값만 사용한다.&lt;/li&gt;&lt;li&gt;이를 히트맵으로 표현하면 Grad-CAM이 된다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Grad-CAM과 Guided Backprop을 결합하여 사용할 수도 있다. Grad-CAM은 rough하고 smooth한 형태를 가지고 있고, Guided Backprop은 sharp하지만 class 구분성이 조금 떨어지므로, 이 두 개를 결합한 Guided Grad-CAM을 이용하면 해당 클래스에 대해 명확한 인식을 가지면서도 sharp하게 모양을 잡아낼 수 있다.&lt;/p&gt;&lt;h3 id=&quot;scouter&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#scouter&quot; aria-label=&quot;scouter permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;SCOUTER&lt;/h3&gt;&lt;p&gt;최근에는 Grad-CAM을 좀 더 개선해서, &amp;quot;이 영상을 무엇으로 판단했느냐&amp;quot; 뿐만 아니라 &amp;quot;이 영상을 왜 그렇게 판단했느냐&amp;quot;까지 비교대조해볼 수 있는 SCOUTER 방법도 제안되었다.&lt;/p&gt;&lt;hr/&gt;&lt;h3 id=&quot;reference&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#reference&quot; aria-label=&quot;reference permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Reference&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;http://karthink.me/journal/class-activation-map.html&quot;&gt;Class Activation Map&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://www.google.com/url?sa=i&amp;amp;url=https%3A%2F%2Fmedium.com%2F%40mohamedchetoui%2Fgrad-cam-gradient-weighted-class-activation-mapping-ffd72742243a&amp;amp;psig=AOvVaw3libP0Oxkiv-LXBqXaTgAI&amp;amp;ust=1615482875580000&amp;amp;source=images&amp;amp;cd=vfe&amp;amp;ved=0CA0QjhxqFwoTCJD30-mcpu8CFQAAAAAdAAAAABAD&quot;&gt;Grad-CAM- Gradient-weighted Class Activation Mapping&lt;/a&gt;&lt;/p&gt;&lt;hr/&gt;&lt;p&gt;개인적으로 자연어처리 LSTM 이후 역대급으로 어려웠던 파트인 것 같다. 너무 많은 내용들이 나오고 방식도 생소해 제대로 이해하고 적은게 반도 안되는 듯하다. 재공부와 필기보충이 필요하다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[Computer Vision 03 - Semantic segmentation]]></title><description><![CDATA[Semantic segmentation by 오태현 교수님, BoostCamp AI Tech 7주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/39_semantic_segmentation/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/39_semantic_segmentation/</guid><pubDate>Tue, 09 Mar 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;semantic-segmentation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#semantic-segmentation&quot; aria-label=&quot;semantic segmentation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Semantic segmentation&lt;/h1&gt;&lt;p&gt;이미지 분류를 영상 단위가 아니라 픽셀 별로 하는 것. 단, 같은 클래스(종류)이면서 서로 다른 물체(개체)를 구분하지는 않는다.(Don&amp;#x27;t care about instances. Only care about &lt;strong&gt;semantic category&lt;/strong&gt;.)&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;적용-분야&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A0%81%EC%9A%A9-%EB%B6%84%EC%95%BC&quot; aria-label=&quot;적용 분야 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;적용 분야&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;의료 이미지&lt;/li&gt;&lt;li&gt;자율 주행&lt;/li&gt;&lt;li&gt;computational photography(object별로 조작할 수 있는 사진) 등&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;object의 구별이 쉬워지므로, object 별 이미지 수정을 하기 위한 인터페이스를 만드는 post-processing에 사용되기도 한다.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;semantic-segmentation-neural-net의-종류&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#semantic-segmentation-neural-net%EC%9D%98-%EC%A2%85%EB%A5%98&quot; aria-label=&quot;semantic segmentation neural net의 종류 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Semantic segmentation Neural Net의 종류&lt;/h2&gt;&lt;h3 id=&quot;fully-convolutional-networksfcn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#fully-convolutional-networksfcn&quot; aria-label=&quot;fully convolutional networksfcn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Fully Convolutional Networks(FCN)&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;최초의 end-to-end semantic segmentation NN&lt;ul&gt;&lt;li&gt;입력과 출력 페어만 있으면 신경망 내부가 자동으로 학습되는 구조&lt;ul&gt;&lt;li&gt;이전까지는 내부 알고리즘을 직접 작성하고 결합하여 만들었었다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;학습시 사용했던 이미지와 입력 이미지의 resolution이 달라도 문제없이 작동한다(호환성이 높다).&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;기존의 CNN은 마지막 부분에 FC layer를 몇 단 두었었는데, FCN은 FC 대신 Fully convolutional layer만 사용한다. 이러한 방식이 어떤 차이가 있을까?&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:37.5%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;fully-connected-vs-fully-convolutional&quot; title=&quot;fully-connected-vs-fully-convolutional&quot; src=&quot;/static/fca0206fac05348d0205787b9d8efacd/2bef9/fully-connected-vs-fully-convolutional.png&quot; srcSet=&quot;/static/fca0206fac05348d0205787b9d8efacd/6f3f2/fully-connected-vs-fully-convolutional.png 256w,/static/fca0206fac05348d0205787b9d8efacd/01e7c/fully-connected-vs-fully-convolutional.png 512w,/static/fca0206fac05348d0205787b9d8efacd/2bef9/fully-connected-vs-fully-convolutional.png 1024w,/static/fca0206fac05348d0205787b9d8efacd/71c1d/fully-connected-vs-fully-convolutional.png 1536w,/static/fca0206fac05348d0205787b9d8efacd/a878e/fully-connected-vs-fully-convolutional.png 2048w,/static/fca0206fac05348d0205787b9d8efacd/42267/fully-connected-vs-fully-convolutional.png 2670w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;em&gt;Fully &lt;strong&gt;connected&lt;/strong&gt; layer&lt;/em&gt; : 공간 정보를 고려하지 않고, fixed vector가 input으로 주어지면, output도 fixed vector로 처리된다.&lt;/li&gt;&lt;li&gt;&lt;em&gt;Fully &lt;strong&gt;convolutional&lt;/strong&gt; layer&lt;/em&gt; : 입/출력이 모두 activation map(tensor)이다. 1x1 conv layer이다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:32.03125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA9klEQVQY042QO0/DMBSF+29BomLowB9h6NQKNiqBEKqoOhQ2HkMVikSSyRXBTeS0qe3Edp1X7V6IqkoscIeju5zznXtb1lpjDGhVlVJKIVRdb+3/ptXY4pgNR6h36d/do0zk+0Tzh/mHWt7cktNO0DlD593pfP7luR9vM//l1UMoYazmqeZcS1mVZaV1rnUBCgW/yUWR9S+WRyf8uI0G106yTihNZu+L8TiYPIRPz7HnEtcljhMIsVmt6CKMASClasjW89XVIJg8riHxV7c814SEEIfxJ8Y4iiLG6OHm/cM21m5hNYeB3UI9xniaZoBi4KNUKdWYdwi1TeZaEVo7AAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;fully-connected&quot; title=&quot;fully-connected&quot; src=&quot;/static/07f47692aa270057fbc6a5c4de8e5efc/2bef9/fully-connected.png&quot; srcSet=&quot;/static/07f47692aa270057fbc6a5c4de8e5efc/6f3f2/fully-connected.png 256w,/static/07f47692aa270057fbc6a5c4de8e5efc/01e7c/fully-connected.png 512w,/static/07f47692aa270057fbc6a5c4de8e5efc/2bef9/fully-connected.png 1024w,/static/07f47692aa270057fbc6a5c4de8e5efc/71c1d/fully-connected.png 1536w,/static/07f47692aa270057fbc6a5c4de8e5efc/df438/fully-connected.png 1556w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;Fully connected layer는 각 채널들을 일직선으로 쭉 펴서(flatten) concat한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:72.265625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;fully-convolutional&quot; title=&quot;fully-convolutional&quot; src=&quot;/static/c1e5b12d3dc8e8390acca8ad40c77d3b/2bef9/fully-convolutional.png&quot; srcSet=&quot;/static/c1e5b12d3dc8e8390acca8ad40c77d3b/6f3f2/fully-convolutional.png 256w,/static/c1e5b12d3dc8e8390acca8ad40c77d3b/01e7c/fully-convolutional.png 512w,/static/c1e5b12d3dc8e8390acca8ad40c77d3b/2bef9/fully-convolutional.png 1024w,/static/c1e5b12d3dc8e8390acca8ad40c77d3b/71c1d/fully-convolutional.png 1536w,/static/c1e5b12d3dc8e8390acca8ad40c77d3b/c655d/fully-convolutional.png 1586w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;이와 달리, Fully convolutional layer는 각 채널에서 같은 feature로 분류되는 vector(즉, 같은 위치의 벡터)들을 묶어 m개의 1x1 필터와 conv 연산(내적)을 수행하여, m개의 채널을 가지는 feature map을 구성한다. 이 때, conv 연산이므로 sliding window 방식을 사용하기 때문에 좀 더 spatial한 데이터가 유지된다는 장점이 있다.&lt;/p&gt;&lt;p&gt;이 과정에서 Pooling 계층을 여러번 통과하고, stride가 있어 FC layer에 비해 좀 더 넓은 receptive field를 가지고 있기 때문에, high-resolution의 input이 들어오더라도 결과값은 &lt;strong&gt;훨씬 low-resolution의 예측 스코어(히트맵)&lt;/strong&gt;를 가지게 된다. &lt;/p&gt;&lt;p&gt;그렇다면 Pooling과 Stride를 제거해버리면 어떨까? 만약 그렇게 한다면 receptive field 자체가 작아져서, 영상의 전체적인 context를 파악하지 못하게 될 것이다. 그건 의미가 없다.&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;upsampling&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#upsampling&quot; aria-label=&quot;upsampling permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Upsampling&lt;/h3&gt;&lt;p&gt;따라서, 이런 저해상도 문제를 피하기 위해 &lt;strong&gt;&lt;code&gt;upsampling&lt;/code&gt;&lt;/strong&gt;을 사용한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:54.6875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;upsampling&quot; title=&quot;upsampling&quot; src=&quot;/static/d9b3edab59a9851e796b241688e561f2/2bef9/upsampling.png&quot; srcSet=&quot;/static/d9b3edab59a9851e796b241688e561f2/6f3f2/upsampling.png 256w,/static/d9b3edab59a9851e796b241688e561f2/01e7c/upsampling.png 512w,/static/d9b3edab59a9851e796b241688e561f2/2bef9/upsampling.png 1024w,/static/d9b3edab59a9851e796b241688e561f2/71c1d/upsampling.png 1536w,/static/d9b3edab59a9851e796b241688e561f2/913b9/upsampling.png 1822w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;고해상도 이미지 input에 Conv 연산과 Pooling을 거치면서, 출력값은 자연스럽게 저해상도로 Downsampling된다. 이를 마지막에 고해상도로 키워주는 연산이 Upsampling이다. 방법은 여러가지가 있지만, 최근에는 대표적으로 2가지가 사용된다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:56.640625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;transposed-convolution&quot; title=&quot;transposed-convolution&quot; src=&quot;/static/1cb98ee87e0384211730d2a03137ce21/2bef9/transposed-convolution.png&quot; srcSet=&quot;/static/1cb98ee87e0384211730d2a03137ce21/6f3f2/transposed-convolution.png 256w,/static/1cb98ee87e0384211730d2a03137ce21/01e7c/transposed-convolution.png 512w,/static/1cb98ee87e0384211730d2a03137ce21/2bef9/transposed-convolution.png 1024w,/static/1cb98ee87e0384211730d2a03137ce21/ccf0c/transposed-convolution.png 1428w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;code&gt;Transposed convolution&lt;/code&gt; : swapping the forward and backward passes of convolution&lt;/p&gt;&lt;p&gt;input 픽셀 각각을 filter에 곱해주어 늘리는 것이다.&lt;/p&gt;&lt;p&gt;잘 이해가 가지 않는다면 다음 링크를 참조해보도록 하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://gaussian37.github.io/dl-concept-transposed_convolution/&quot;&gt;Transposed Convolution을 이용한 Upsampling&lt;/a&gt;&lt;/p&gt;&lt;p&gt;이 때 반드시 주의해야 할 점은, &lt;strong&gt;&lt;em&gt;checkboard artifact&lt;/em&gt;&lt;/strong&gt;를 조심해야한다는 것이다. 마치 체크무늬처럼 일정 간격으로 특정 격자의 색이 진하게 나오는 것을 의미한다. 이는 &lt;strong&gt;&lt;div&gt;sliding window 기법을 사용하는 CNN의 특성 상 stride와 kernel 크기를 잘 조절하지 않으면 overlap 되는 구간이 생기기 때문에 일어난다.&lt;/div&gt;&lt;/strong&gt;&lt;/p&gt;&lt;p&gt;위의 이미지에서도 볼수 있듯, [az + bx] 구간처럼 overlap되는 구간은 다른 구간들보다 상대적으로 출력값이 높을 수 밖에 없다. 픽셀값이 높다는 것은 곧 진한 색상을 의미하므로, 상대적으로 진한 격자가 규칙적으로 나타나게 되는 것이다. 이는 짙은 색상의 이미지일수록 더 심하다.&lt;/p&gt;&lt;p&gt;&lt;code&gt;Upsample and convolution&lt;/code&gt; : Decompose into spatial upsampling and featrue convolutoin&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:18.75%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAECAIAAAABPYjBAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA3ElEQVQI1yWKS2vCQBSF8/frg+606cMHrqWoTSvtwoipG1EbZzKZiRloM7djQtVYNyKKetXD4cD3cQyPeQpUksTJOXF6CcCP1nAtgEKzTJcAkdYKze9ZRovF3Hj/+Oo5/Fq7y0I5IyQqmna5+vlUcUoVx3ywGVfu5LtgdhAfSz3cu/vOcCSNaq391h6+WAPrdVB/7iv4EyLO5JqZXCObb1y2FcpEBChbN1nEJm7+1mJMG0JwSidc+L7PsNvtZr/fyRCZBAGj1JUhPx4Pq1VKyHg6xRsRwvOou17/nwDG6sBDx8EfKgAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;upsample-and-convolution&quot; title=&quot;upsample-and-convolution&quot; src=&quot;/static/213d8f774fd5f3226cb14564e5d26c1f/2bef9/upsample-and-convolution.png&quot; srcSet=&quot;/static/213d8f774fd5f3226cb14564e5d26c1f/6f3f2/upsample-and-convolution.png 256w,/static/213d8f774fd5f3226cb14564e5d26c1f/01e7c/upsample-and-convolution.png 512w,/static/213d8f774fd5f3226cb14564e5d26c1f/2bef9/upsample-and-convolution.png 1024w,/static/213d8f774fd5f3226cb14564e5d26c1f/71c1d/upsample-and-convolution.png 1536w,/static/213d8f774fd5f3226cb14564e5d26c1f/a878e/upsample-and-convolution.png 2048w,/static/213d8f774fd5f3226cb14564e5d26c1f/b297c/upsample-and-convolution.png 2844w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;이러한 문제는 upsampling(정확히는 interpolution) 과정과 convolution 과정을 분리함으로써 쉽게 해결할 수 있다. Transposed convolution은 어설프게 overlap되는 구간들이 일부만 있었다면, upsampling을 통해 중첩 문제가 없이 골고루 영향을 받게 함으로써 전체적으로 평준화시켜줄 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;기존의 upsampling 연산 decomposition → &lt;strong&gt;[spatial upsampling + feature convolution]&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;spacial upsampling&lt;/strong&gt; : {Nearest-neighbor (NN), Bilinear sampling}같은 interpolation. 학습가능한 파라미터는 전혀 없고 그냥 해상도를 키워주는 작업.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;convolution 연산&lt;/strong&gt; : spatial upsampling 직후 수행하며, 파라미터가 있으므로 학습이 되어 전체적으로 learnable upsampling을 구현할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;다시-fcn으로&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%8B%A4%EC%8B%9C-fcn%EC%9C%BC%EB%A1%9C&quot; aria-label=&quot;다시 fcn으로 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;다시, FCN으로&lt;/h3&gt;&lt;p&gt;그러나 아무리 Upsampling을 했다고 하더라도 한 번 떨어진 resolution을 충분히 원래대로 끌어올리기는 쉽지 않다. 연산 과정에서 잃어버린 정보들을 다시 살리는 일이기 때문이다.&lt;/p&gt;&lt;p&gt;결국, 목표대로 정확한 High-resolution의 output을 얻기 위해서는 두 마리 토끼를 다 잡는 미션을 수행해야 한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;Fine/Low-level/Detail/Local, 미세한 각 부분의 디테일을 살리면서도&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Coarse/Semantic/Holistic/Global, 전체적인 context를 볼 수 있는 넓은 시야를 가져야한다.&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 두 특징을 모두 가지기위해서 기존의 방법들을 모두 fusion한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:49.21875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;FCN-skip-connection&quot; title=&quot;FCN-skip-connection&quot; src=&quot;/static/caa0e2cc00feeb6bd22b190592b19049/2bef9/FCN-skip-connection.png&quot; srcSet=&quot;/static/caa0e2cc00feeb6bd22b190592b19049/6f3f2/FCN-skip-connection.png 256w,/static/caa0e2cc00feeb6bd22b190592b19049/01e7c/FCN-skip-connection.png 512w,/static/caa0e2cc00feeb6bd22b190592b19049/2bef9/FCN-skip-connection.png 1024w,/static/caa0e2cc00feeb6bd22b190592b19049/71c1d/FCN-skip-connection.png 1536w,/static/caa0e2cc00feeb6bd22b190592b19049/a878e/FCN-skip-connection.png 2048w,/static/caa0e2cc00feeb6bd22b190592b19049/ae3b6/FCN-skip-connection.png 2836w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ol&gt;&lt;li&gt;높은 layer의 activation map을 upsampling하여 해상도를 크게 끌어올린다.&lt;/li&gt;&lt;li&gt;이에 맞추어 중간 layer의 activation map을 upsampling하여 가져오고, concat한다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;위에서는 FCN-8s가 가장 많은 layer들을 concat하는 형태가 된다. 이처럼 중간 layer들의 skip connection을 추가할 때 훨씬 더 명확한 이미지를 얻을 수 있다.&lt;/p&gt;&lt;h3 id=&quot;fcn의-특징&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#fcn%EC%9D%98-%ED%8A%B9%EC%A7%95&quot; aria-label=&quot;fcn의 특징 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;FCN의 특징&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;Faster&lt;ul&gt;&lt;li&gt;직접 짠 컴포넌트(알고리즘)에 의존하지 않고 자동적으로 학습하는 end-to-end 구조이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Accurate&lt;ul&gt;&lt;li&gt;feature 표현과 분류가 함께 최적화된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;hypercolumn-for-object-segmentation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#hypercolumn-for-object-segmentation&quot; aria-label=&quot;hypercolumn for object segmentation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Hypercolumn for object segmentation&lt;/h3&gt;&lt;p&gt;FCN과 동일한 시기에 비슷한 내용의 연구도 있었는데, 이 경우 Fully convolutional network보다는 &lt;code&gt;Hypercolumn&lt;/code&gt;을 강조했다.&lt;/p&gt;&lt;p&gt;기존의 CNN layer는 feature representation을 위해 마지막 FC layer의 출력을 사용했다. 그러나 이 방식은 한 픽셀에 모든 정보가 압축되어 있어 너무 coarse spatially했다.&lt;/p&gt;&lt;p&gt;이와 달리, Hypercolumn은 모든 CNN 유닛의 해당 픽셀 위치에 해당하는 값들을 stacked vector로 표현하는 방식이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;초기 layer에서의 좀 더 미세한 국지적 정보들(fine localized information)을 추출할 수 있었다.&lt;/li&gt;&lt;li&gt;후기 layer에서는 더 전체적인 context를 보므로, coarse semantic information을 추출할 수 있었다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;다만 이 논문은 end-to-end가 아니라, bounding box를 구하는 sub component algorithm을 사용한 뒤 적용하는 모델로 소개되었다.&lt;/p&gt;&lt;h2 id=&quot;u-net&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#u-net&quot; aria-label=&quot;u net permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;U-Net&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;FCN을 베이스로 만들어졌다.&lt;/li&gt;&lt;li&gt;낮은 층의 feature map과 높은 층의 feature map을 더 잘 결합하는 방식을 제시했다.&lt;ul&gt;&lt;li&gt;FCN의 skip connection과 유사한 방식이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;좀 더 정교한 segmentation이 가능해졌다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;구조&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B5%AC%EC%A1%B0&quot; aria-label=&quot;구조 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;구조&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:44.53125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;u-net-architecture&quot; title=&quot;u-net-architecture&quot; src=&quot;/static/c9af1e7fc0bb5189c67506ace52a3926/2bef9/u-net-architecture.png&quot; srcSet=&quot;/static/c9af1e7fc0bb5189c67506ace52a3926/6f3f2/u-net-architecture.png 256w,/static/c9af1e7fc0bb5189c67506ace52a3926/01e7c/u-net-architecture.png 512w,/static/c9af1e7fc0bb5189c67506ace52a3926/2bef9/u-net-architecture.png 1024w,/static/c9af1e7fc0bb5189c67506ace52a3926/71c1d/u-net-architecture.png 1536w,/static/c9af1e7fc0bb5189c67506ace52a3926/a878e/u-net-architecture.png 2048w,/static/c9af1e7fc0bb5189c67506ace52a3926/ea66f/u-net-architecture.png 2828w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;기존의 CNN 파트와 conv 연산을 적용하여 전체적인 feature map(holistic context)을 뽑아내는 downsampling 부분은 거의 같다. 여기서는 &lt;code&gt;Contracting path&lt;/code&gt;라고 부른다. &lt;/p&gt;&lt;ul&gt;&lt;li&gt;3x3 conv를 사용한다.&lt;/li&gt;&lt;li&gt;feature 채널의 숫자를 계속 doubling한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;그러나, upsampling 파트는 차이가 있다. 여기에서는 &lt;code&gt;Expanding path&lt;/code&gt;, &lt;code&gt;Decoding&lt;/code&gt; 이라고 부른다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;한번에 upsampling 하는 대신, 채널 수를 반으로 줄여가며 점진적으로 upsampling한다(즉, Contracting path의 대응되는 layer와 채널 수를 동일하게 맞춘다.).&lt;/li&gt;&lt;li&gt;대칭되는 Contracting path의 layer에서 skip connection을 통해 대칭되는 feature map들을 가져와서 fusion(여기서는 concat)해준다.&lt;/li&gt;&lt;/ul&gt;&lt;div&gt;&lt;p&gt;이 때 input 이미지와 feature 이미지의 크기는 짝수여야 한다. 만약 홀수라면, Contracting/Expanding 파트에서 나머지 정보들이 유실된다.&lt;/p&gt;&lt;/div&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;deeplab&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#deeplab&quot; aria-label=&quot;deeplab permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;DeepLab&lt;/h2&gt;&lt;p&gt;Sementic segmentation의 한 획을 그은 모델로, 기존 모델에 비해 다음과 같은 차이점을 가진다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;em&gt;CRFs&lt;/em&gt;&lt;/strong&gt; 후처리의 존재&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Atrous convolution&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;conditional-random-fieldscrfs&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#conditional-random-fieldscrfs&quot; aria-label=&quot;conditional random fieldscrfs permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Conditional Random Fields(CRFs)&lt;/h3&gt;&lt;p&gt;픽셀과 픽셀 사이의 관계를 다 이어주고, regular한 pixel map을 그리드로 본다.&lt;/p&gt;&lt;p&gt;기존의 sementic segmentation에서는 feedforward 구조이므로 피드백이 없어 굉장히 blurry한 output이 나오기 마련인데, CRFs는 기존의 이미지에서 edge같은 경계선들을 활용하여 score map이 경계에 잘 들어맞도록 확산시켜주는 역할을 한다. 이 때 물체의 background와 내부에서 동시에 확산하므로, 결과적으로 경계선이 물체 형태에 맞게 명확히 잡히게 된다.&lt;/p&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;atrous-convolution&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#atrous-convolution&quot; aria-label=&quot;atrous convolution permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Atrous convolution&lt;/h3&gt;&lt;p&gt;기존 convolution 필터와 달리, 필터의 수용영역 사이사이에 space를 넣어 spatial context를 캐치하는 방법이다. Dilation factor를 몇번 반복하는 것 만으로 파라미터 수는 늘리지 않으면서 receptive field는 exponential하게 키울 수 있다.&lt;/p&gt;&lt;h1 id=&quot;-7&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-7&quot; aria-label=&quot; 7 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;depthwise-separable-convolution&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#depthwise-separable-convolution&quot; aria-label=&quot;depthwise separable convolution permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Depthwise separable convolution&lt;/h3&gt;&lt;p&gt;기존의 convolution 연산은 하나의 필터를 모든 input 채널에 대입시켰다. &lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:42.96875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAJCAIAAAC9o5sfAAAACXBIWXMAABYlAAAWJQFJUiTwAAABQElEQVQoz41RW1KEMBDk/vdSb+DHugphE8gbEkLIA9RmOYBOFalJM91ND02McRjGtm2ttSkl59y+7z//qwYP+I8HhQQhPcoYq5Q2xkDoDzLmSUcYY4SQEMJxHN/Pgr8Usuu6cRxxYoBzgQ+E08XEZKOl5EJu21Zy0caGsM7rAmupTMq51uKcX0LIJXvvMHka7Iex0zTNzf31raUD7dnA4KRjyvfb++3lrXswOowDfSCA0pEy9XX7+Gx7yjAsuNAQbayZjJ2XZTFm0maCc1iClBp9jBtAeJA+sCHO3ittw7oClNpOs2twX1cEiVwoACllB7I2SJELgpzSpRTvFwSZna+1XtLblprrx9S644Jmf5bWWIVUSnWkRwMcfCFOBFkppWiw1KY8C384n+up2BwaIJC4No/zIl+KQgjOufcer34BPn39yLdOwTsAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;depthwise convolution&quot; title=&quot;depthwise convolution&quot; src=&quot;/static/3f3ecade6a44e1eff0b562b907a7d1e5/2bef9/depthwise-covolution.png&quot; srcSet=&quot;/static/3f3ecade6a44e1eff0b562b907a7d1e5/6f3f2/depthwise-covolution.png 256w,/static/3f3ecade6a44e1eff0b562b907a7d1e5/01e7c/depthwise-covolution.png 512w,/static/3f3ecade6a44e1eff0b562b907a7d1e5/2bef9/depthwise-covolution.png 1024w,/static/3f3ecade6a44e1eff0b562b907a7d1e5/71c1d/depthwise-covolution.png 1536w,/static/3f3ecade6a44e1eff0b562b907a7d1e5/a878e/depthwise-covolution.png 2048w,/static/3f3ecade6a44e1eff0b562b907a7d1e5/1105b/depthwise-covolution.png 2940w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;그러나 &lt;em&gt;Depthwise separable convolution&lt;/em&gt;은 이 과정을 둘로 분리하였다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;Depthwise convolution : 채널별로 필터를 만들고, 이를 convolution하여 채널별로 activation map을 만든다.&lt;/li&gt;&lt;li&gt;Pointwise convolution : 뽑아낸 값들을 토대로 다시 1x1 conv를 사용함으로써 하나의 값으로 출력이 되도록 만들어준다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;이미지를 보면 알 수 있겠지만, convolution의 표현력은 유지하면서 계산량은 훨씬 더 줄어들게 되었다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[Computer Vision 01 - 이미지 분류(Image Classification)]]></title><description><![CDATA[Image Classification by 오태현 교수님, BoostCamp AI Tech 7주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/37_image_classification/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/37_image_classification/</guid><pubDate>Mon, 08 Mar 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;blockquote&gt;&lt;p&gt;&lt;em&gt;The theory and development of computer systems able to perform tasks that normally require human intelligence, such as visual perception, speech recognition, decision-making, and translation between languages. - from oxford dictionary&lt;/em&gt;&lt;/p&gt;&lt;/blockquote&gt;&lt;p&gt;인간의 지능을 구현하기 위해서는 가장 먼저 &lt;strong&gt;다양한 지각능력(multi-modal perception)&lt;/strong&gt;을 구현해야한다. 지각능력은 입력(input)과 출력(output)을 통해 구현되며, 이 때 입력은 단순히 오감에서 끝나는 것이 아니라, face/touch/speech/social같은 복합적인 감각의 이해도 포함된다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;시각적 지각능력(visual perception)&lt;/strong&gt;은 인간 지능의 한 부분으로 취급된다. &lt;strong&gt;&lt;code&gt;컴퓨터 비젼(Computer Vision)&lt;/code&gt;&lt;/strong&gt;은 이러한 지각능력을 재현하고자 하는 분야이다. 시각의 프로세스는 &lt;em&gt;[Visual world - Sensing device(눈) - Interpreting device(뇌) - Interpretation]&lt;/em&gt; 네 단계로 이루어진다.  마찬가지로, 컴퓨터 비젼도 동일한 네 단계로 이루어진다. 다만 Sensing device가 눈이 아닌 카메라이고, Interpreting device가 뇌가 아닌 GPU와 알고리즘이라는 것만 다르다. 이를 통해 나오게 된 해석(Interpretation, representation)은 하이 레벨의 description이 된다.  이 representation을 통해 장면에 해당하는 이미지나 3D 모델을 재구현하는것을 &lt;code&gt;Computer Graphics&lt;/code&gt;, 또는 &lt;code&gt;렌더링(Rendering)&lt;/code&gt;이라고 한다. 거꾸로, 시각적 데이터(visual data)에서 representation을 추출하는 일을 &lt;strong&gt;&lt;code&gt;Inverse Rendering&lt;/code&gt;&lt;/strong&gt;이라고 하며, 컴퓨터 비젼의 task에 해당한다.&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h1 id=&quot;image-classification&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#image-classification&quot; aria-label=&quot;image classification permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Image Classification&lt;/h1&gt;&lt;p&gt;분류기(Classifier)는 입력을 어떤 카테고리 값과 매핑시켜 내보내는 장치이다. 이미지 분류는 이 분류기의 입력값으로 시각적 데이터만을 사용하여 추론하는 것을 일컫는다.&lt;/p&gt;&lt;p&gt;극단적으로 생각해보았을 때, 모든 분류 문제는 세상의 모든 시각적 데이터를 가지고 있다면 아주 쉽게 해결된다. 그냥 모든 데이터들 사이에서 비슷한 것들끼리 모으기만 하면 된다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:600px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:50%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAKCAYAAAC0VX7mAAAACXBIWXMAAAsSAAALEgHS3X78AAACpElEQVQoz2NggALlSfWcDPUZorwdJcK6U+vZwIJZqRDJ3EAGZGA4vZWNva1AiLejWFR9ci3X////GRgkGRjkVh9mYDCZ0V6kNqnBWm5i3TGNqc3fNKY2fZOdWLeeobGcFaRZdlIjI9gULaBBM9ur9XprdcR6K+eqTWn6DlT/Vaq/5rJgT6UGSInShEpGBrOsRD/TpCg59QU98jrT2hwUp7ZZhs6uUj03I5gFpMh/VRVDhJI22Eyj5lI302A/C6ENUwV1pzbbyU9utnWfUanfMDOT6/8iqPPzzB34Kw2dWGDeAbqe+f9ZMcn/xyWFQPx/i4wYi2rj4N5NCggWQVLL9P8Sv9j/E2LCcAWBoWHi4TkOrHoZ2zgdHBZI/9hlKXStKsX4dnGw3uWUbPaba+2Yr/V5M4HUeiXEcgQGBop5/7/B7DihRfLuQhP+a7WZhnfKw4wux/nxgQ0M8fQVjVvYxh/y4IBgRttkxQPp6UKXkrIlLySlSDKggXBHN7aI9CSpqP8nBTMW9cpPic8WvJldxHcqIkPoZHAYO1hRqKuXsNOheYI+3w7oFpyZI/UwLEb5Wk6q1in7ePGvayItr87w63q1JKQYpDZCyZA1qLlYzv/TfiOj/8f4b8Qn6V7NzfW5mpLsdDclmQdsYLCdi1jVtCad6nUTtfOqcriRXfR/ib/Rh1lexS+XRSaC+OoMPKwZDUXGtRsnG+U2FPAejU+UOJOWFnA8LU0frik4OFh1RkqSy7qoGPv6wGCFwLAwsbDQMKnZXt58EvFxQQxd3W4wtbbBwXzVsTFOG+PjXaaFRegEh4aIx4WGSsX7BQo4hodYGLRX5jFU19awoLgKGHUVT86D057B1EZTy1ltCSC25YPjDLP7uhkZ3v1hwAbc5vbxAdNpDABeeelx7c/gjgAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;k-nn&quot; title=&quot;k-nn&quot; src=&quot;/static/5680ba7394c320efaa16adfe9b5a1cf2/0a47e/k-nn.png&quot; srcSet=&quot;/static/5680ba7394c320efaa16adfe9b5a1cf2/6f3f2/k-nn.png 256w,/static/5680ba7394c320efaa16adfe9b5a1cf2/01e7c/k-nn.png 512w,/static/5680ba7394c320efaa16adfe9b5a1cf2/0a47e/k-nn.png 600w&quot; sizes=&quot;(max-width: 600px) 100vw, 600px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;즉, &lt;strong&gt;K Nearest Neighbors(K-NN) 문제&lt;/strong&gt;로 해결할 수 있다. K-NN 문제는 단순히 이미지 레이블 데이터 값을 주위의 다른 데이터 레이블들과 비교하여 가장 비슷하다고 판단되는 후보군으로 편입시키는 문제이다. 이렇게 해결하는 분류기가 있다면, 마치 검색엔진처럼 작동한다. 그러나, 이러한 접근 방식은 불가능하다. Time/Memory Complexity 무한대일 것이라는 점과, &amp;#x27;비슷하다&amp;#x27;는 기준을 어떻게 잡을건지가 모호하다는 것이 결정적인 불가능 요인이다. 따라서 컴퓨터 비젼은 방대한 데이터를 제한된 complexity의 시스템(인공 신경망)이라는 분류기에 녹여넣는 것이 목표이다.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;fully-connected-layer-network&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#fully-connected-layer-network&quot; aria-label=&quot;fully connected layer network permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Fully Connected Layer Network&lt;/h3&gt;&lt;p&gt;이런 이미지 분류를 가장 간단한 형태의 인공 신경망 분류기, 즉 &lt;strong&gt;단일 계층의 Fully Connected Layer Network&lt;/strong&gt;로 구현했다고 생각해보자.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:40.625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;fully-connected-layer-network&quot; title=&quot;fully-connected-layer-network&quot; src=&quot;/static/9c28f194fefe28a0fad68d5d88ec2be5/2bef9/fully-connected-layer-network.png&quot; srcSet=&quot;/static/9c28f194fefe28a0fad68d5d88ec2be5/6f3f2/fully-connected-layer-network.png 256w,/static/9c28f194fefe28a0fad68d5d88ec2be5/01e7c/fully-connected-layer-network.png 512w,/static/9c28f194fefe28a0fad68d5d88ec2be5/2bef9/fully-connected-layer-network.png 1024w,/static/9c28f194fefe28a0fad68d5d88ec2be5/71c1d/fully-connected-layer-network.png 1536w,/static/9c28f194fefe28a0fad68d5d88ec2be5/a878e/fully-connected-layer-network.png 2048w,/static/9c28f194fefe28a0fad68d5d88ec2be5/30faf/fully-connected-layer-network.png 2802w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;기존의 이미지에 W(weight)를 내적한 것을 다시 이미지로 치환시켜보면, 아래와 같이 실제 레이블과 어느정도 관련이 있는 이미지 형상을 추출할 수 있다. 그러나 이 형태는 너무 흐릿하여 정확도가 떨어진다. &lt;strong&gt;계층을 너무 얕게 쌓았기 때문에 디테일을 충분히 표현할 수 없는 것&lt;/strong&gt;이다. &lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:67.578125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;cropped-fully-connected-layer-network&quot; title=&quot;cropped-fully-connected-layer-network&quot; src=&quot;/static/69fce1d11ae55b460b98d0c216c66f99/2bef9/cropped-fully-connected-layer-network.png&quot; srcSet=&quot;/static/69fce1d11ae55b460b98d0c216c66f99/6f3f2/cropped-fully-connected-layer-network.png 256w,/static/69fce1d11ae55b460b98d0c216c66f99/01e7c/cropped-fully-connected-layer-network.png 512w,/static/69fce1d11ae55b460b98d0c216c66f99/2bef9/cropped-fully-connected-layer-network.png 1024w,/static/69fce1d11ae55b460b98d0c216c66f99/71c1d/cropped-fully-connected-layer-network.png 1536w,/static/69fce1d11ae55b460b98d0c216c66f99/2aa89/cropped-fully-connected-layer-network.png 1698w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;또, 모든 노드들이 출력층으로 전연결(fully-connect)되기 때문에, &lt;strong&gt;전체적인 형상만을 학습&lt;/strong&gt;한다. 따라서 만약 이 사진을 잘라낸 사진을 제공하면 인식을 전혀 할 수 없게 된다.&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;convolutional-neural-networkcnn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#convolutional-neural-networkcnn&quot; aria-label=&quot;convolutional neural networkcnn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Convolutional Neural Network(CNN)&lt;/h2&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:686px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:44.140625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;fully-vs-locally&quot; title=&quot;fully-vs-locally&quot; src=&quot;/static/ec56970b36992c0cc8866dfdaa8965d7/f6386/fully-vs-locally.png&quot; srcSet=&quot;/static/ec56970b36992c0cc8866dfdaa8965d7/6f3f2/fully-vs-locally.png 256w,/static/ec56970b36992c0cc8866dfdaa8965d7/01e7c/fully-vs-locally.png 512w,/static/ec56970b36992c0cc8866dfdaa8965d7/f6386/fully-vs-locally.png 686w&quot; sizes=&quot;(max-width: 686px) 100vw, 686px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;이러한 전연결 신경망의 한계점을 극복하기 위하여 &lt;strong&gt;&lt;code&gt;합성곱 신경망(Convolution Neural Network, CNN)&lt;/code&gt;&lt;/strong&gt;가 나오게 되었다. CNN은 모든 노드들을 다음 계층으로 전연결시키는 것이 아니라, &lt;strong&gt;국소적인 연결(locally connect)&lt;/strong&gt;을 사용한다. 동일한 국소적 sliding window를 이미지의 모든 부분에 대입시켜 feature들을 뽑아냄으로써, 치우쳐 있는 이미지나 잘린 이미지라도 feature를 추출할 수 있고, 파라미터를 재활용하여 메모리도 적게 사용할 수 있다. 이런 장점 때문에 많은 CV task의 backbone으로 활용되고 있다.&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;이미지-분류를-위한-cnn-아키텍쳐의-종류&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9D%B4%EB%AF%B8%EC%A7%80-%EB%B6%84%EB%A5%98%EB%A5%BC-%EC%9C%84%ED%95%9C-cnn-%EC%95%84%ED%82%A4%ED%85%8D%EC%B3%90%EC%9D%98-%EC%A2%85%EB%A5%98&quot; aria-label=&quot;이미지 분류를 위한 cnn 아키텍쳐의 종류 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;이미지 분류를 위한 CNN 아키텍쳐의 종류&lt;/h2&gt;&lt;h3 id=&quot;간략한-역사-개요&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B0%84%EB%9E%B5%ED%95%9C-%EC%97%AD%EC%82%AC-%EA%B0%9C%EC%9A%94&quot; aria-label=&quot;간략한 역사 개요 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;간략한 역사 개요&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;LeNet&lt;/code&gt;&lt;ul&gt;&lt;li&gt;1998년 Yann Lecun&lt;/li&gt;&lt;li&gt;비교적 간단하며, 한 글자 정도의 손글씨를 인식하기 위해 사용되었다.&lt;/li&gt;&lt;li&gt;구조 : Conv-Pool-Conv-Pool-FC-FC&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;AlexNet&lt;/code&gt;&lt;ul&gt;&lt;li&gt;LeNet에서 모티베이션을 따왔다.&lt;/li&gt;&lt;li&gt;파라미터와 학습 데이터를 훨씬 더 크게 늘렸다.&lt;/li&gt;&lt;li&gt;필터 사이즈가 11x11로 아주 크다. 최근에는 이런 큰 필터를 사용하지 않는다.&lt;/li&gt;&lt;li&gt;활성화 함수로 ReLU를 사용하고, dropout 정규화 기법을 사용했다.&lt;/li&gt;&lt;li&gt;논문에는 메모리 문제로 두 GPU에 올려서 학습했으며, 그 당시 명암을 조정하기 위해 사용했던 LRN(Local Response Normalization) 기법은 현재는 사용하지 않는다.&lt;/li&gt;&lt;li&gt;구조 : Conv  - Pool - LRN - Conv - Pool - LRN - Conv - Conv - Conv - Pool - FC - FC - FC&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;VGGNet&lt;/code&gt;&lt;ul&gt;&lt;li&gt;3x3의 작은 필터와 2x2 max pooling 사용, LRN 제거로 아키텍쳐가 비교적 간단해졌으나 성능은 더 좋아졌다.&lt;/li&gt;&lt;li&gt;19 layer로 AlexNet(12 layer)보다 더 깊다.&lt;ul&gt;&lt;li&gt;작은 필터크기임에도 불구하고, 더 깊이 층을 쌓아 receptive field의 크기를 키웠다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;미리 학습된 feature를 fine-tuning하지 않고도 다른 task에 적용 가능할 정도로 일반화가 잘 되었다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;AlexNet에서 VGGNet으로 발전하면서, 더 깊은 네트워크일수록 더 좋은 성능을 낸다는 것을 확인했다. 그렇다면 과연 층을 단순히 더 깊게 쌓으면, 항상 더 좋은 네트워크를 얻을 수 있을까? 물론 그렇지 않았다. 층을 깊게 쌓으면 쌓을수록 학습을 어렵게 만드는 문제들이 있었다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;기울기 소실/폭발(Gradient vanishing/exploding)&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;연산 복잡도 증가(Computationally complex)&lt;/li&gt;&lt;li&gt;depth가 어느정도 깊어지면, 성능이 떨어지기 시작&lt;ul&gt;&lt;li&gt;&lt;del&gt;Overfitting이 아닐까?&lt;/del&gt; → 기울기 소실/폭발로 인한 Degradation 문제&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이러한 문제점들을 인식한 채로 새로운 네트워크 형태들이 등장하기 시작했다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;GoogLeNet&lt;/code&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;em&gt;Inception Module&lt;/em&gt;&lt;/strong&gt; 을 여러 층 쌓는 형태를 제안한다.&lt;ul&gt;&lt;li&gt;하나의 층에서 다양한 크기의 필터를 사용하여 여러 측면에서 보겠다(depth 확장이 아닌 수평확장)&lt;/li&gt;&lt;li&gt;이 결과들은 모두 concatenation 하여 다음 층으로 넘겨주게 된다.&lt;/li&gt;&lt;li&gt;이 때, 1x1 conv를 한 번 적용해 채널 수를 줄여 계산 복잡도를 떨어뜨린다. 이를 병목(bottleneck) 층이라고 한다.&lt;ul&gt;&lt;li&gt;3x3, 5x5 conv 직전 / pooling 연산 직후&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;em&gt;Auxiliary classifier&lt;/em&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;역전파 중 기울기 소실 문제를 해결하기 위해 Loss를 측정하는 Auxiliary classifier를 네트워크 중간 중간에 둔다.&lt;ul&gt;&lt;li&gt;GoogLeNet 구조 중 Pooling 계층에서 시작해 1x1 conv → FC → FC → softmax로 Loss를 측정하는 구조이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;역전파 시 여기에서도 역전파가 수행되고, 기존의 기울기와 결합하여 학습을 돕는다. 학습이 끝나면 제거한다.&lt;/li&gt;&lt;li&gt;regularization 역할도 수행하는 것으로 알려져있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;ResNet&lt;/code&gt;&lt;ul&gt;&lt;li&gt;최초로 100개가 넘는 layer를 쌓음으로써, 더 깊은 layer를 쌓을수록 성능이 더 좋아진다는 것을 보여준 첫 모델이다. 또한, 인간의 지각 능력을 뛰어넘은 첫 모델이기도 하다.&lt;/li&gt;&lt;li&gt;계기&lt;ul&gt;&lt;li&gt;네트워크 깊이를 늘리다보면 어느 순간부터 정확도(accuracy) 감소가 포화상태(saturated)에 이른다.&lt;/li&gt;&lt;li&gt;기존 인식(가설)&lt;ul&gt;&lt;li&gt;모델 파라미터가 너무 많아지면 overfitting 되어 training error가 더 적고 test error가 더 많은 결과가 나올 것이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;실험 결과&lt;ul&gt;&lt;li&gt;&lt;strong&gt;overfitting 문제가 아니다.&lt;/strong&gt; training error든 test error든 더 깊은 층(56)의 네트워크가 더 얕은 층(20)의 네트워크보다 에러 수가 높게 나온다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;div&gt;degradation 문제이다. 기울기 소실 때문에 최적화(optimization)이 덜 되어 깊은 층의 네트워크가 학습이 덜 된 것이다!&lt;/div&gt;&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;em&gt;residual(skip) connection&lt;/em&gt;&lt;/strong&gt; : 층이 깊어질수록 기존의 input x의 영향력(기울기)이 소실되어 충분히 학습하기 어렵다. 따라서, 몇 개의 층을 지나면 기존의 x와 동일한 값(identity)를 잔차(residual)로 더해주어, 잔여부분만 학습함으로써 학습 부담을 경감시킨다. (분할정복과 비슷한 원리)&lt;ul&gt;&lt;li&gt;역전파 시에도 gradient가 원래 네트워크 레이어 쪽과 skip connection 쪽 두 군데로 흐르므로, 한 곳에서 기울기 소실이 일어나더라도 다른 한쪽을 통해 학습을 정상적으로 지속할 수 있게 된다.&lt;/li&gt;&lt;li&gt;skip connection이 한번 일어날 때마다 역전파 gradient가 흐르는 방법의 경우의 수가 2배로 늘어나므로, 전체 경우의 수는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;2^n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.664392em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.664392em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개가 된다.&lt;/li&gt;&lt;li&gt;residual block은 2개의 3x3 conv layer로 이루어져 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;출력 직전 FC 층은 하나만 존재한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;ResNet의 등장 이후에는 ...&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;DenseNet&lt;/code&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;em&gt;Dense block&lt;/em&gt;&lt;/strong&gt; : x identity를 그대로 더하던 ResNet의 Residual block과 달리, 각 layer의 output을 채널 축 방향으로 concatenation한다.&lt;ul&gt;&lt;li&gt;기울기 소실 문제 해결(Alleviate vanishing gradient problem)&lt;/li&gt;&lt;li&gt;featrue propagation 강화&lt;/li&gt;&lt;li&gt;feature 재활용 가능(Encourage the reuse of features) : concat으로 메모리는 더 많이 쓰게 되지만, 이전 층 layer들의 정보를 합치지 않고 그대로 보존하고 있어 필요할 때 꺼내 쓸 수 있게 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;SENet&lt;/code&gt;&lt;ul&gt;&lt;li&gt;depth를 높이거나 connection을 사용하지 않고, 채널간의 관계를 모델링하여 중요한 특징에 Attention할 수 있도록 만드는 방법.&lt;/li&gt;&lt;li&gt;Attention 생성 방법&lt;ol&gt;&lt;li&gt;&lt;strong&gt;Squeeze&lt;/strong&gt;  : global average pooling으로 각 채널의 공간정보(H,W)를 없애고 분포(또는 magnitude)만 보게 한다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Excitation&lt;/strong&gt; : FC layer를 하나 두어 Squeeze 한 값을 통과시키면서 채널간의 연관성을 나타내는(채널을 reweighting 하는) attention score를 계산한다.&lt;/li&gt;&lt;/ol&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;EfficientNet&lt;/code&gt;&lt;ul&gt;&lt;li&gt;기존에 네트워크 성능을 높이는 방법&lt;ul&gt;&lt;li&gt;deep / wide / (high) resolution scaling&lt;ul&gt;&lt;li&gt;high resolution scaling : 애초에 input 이미지의 resolution이 높으면 성능이 더 좋아진다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;그러나 세 방법은 각각 accuracy saturate의 시점이 다르다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;그래서 세 방법 모두를 적절히 섞은 compound scaling이 등장한다.&lt;/li&gt;&lt;li&gt;지금까지 나왔던 모든 방식들을 대상으로, 적은 FLOP에서도 압도적인 성능차를 보였다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;Deformable Convolution Network&lt;/code&gt;&lt;ul&gt;&lt;li&gt;불규칙적인 2D 이미지에 대한 convolution&lt;ul&gt;&lt;li&gt;사람이나 동물등의 사진은 팔다리가 어떤 형태로 있느냐에 따라 다른 모양을 가진다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;기존의 convolution 필터에 추가로 2D offset map을 추정하기 위한 branch가 따로 결합되어있다.&lt;ul&gt;&lt;li&gt;conv 연산의 결과에 offset field를 추가시키면, offset된 feature map이 나온다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이 모든 모델들은 단순히 말로는 이해하기 어려우니, 꼭 슬라이드와 논문을 참조해가면서 복습하도록 하자.&lt;/p&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;table&gt;&lt;thead&gt;&lt;tr&gt;&lt;th&gt;모델명&lt;/th&gt;&lt;th&gt;구조&lt;/th&gt;&lt;th&gt;메모리요구량&lt;/th&gt;&lt;th&gt;연산량&lt;/th&gt;&lt;th&gt;정확도&lt;/th&gt;&lt;th align=&quot;center&quot;&gt;비고&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;&lt;tbody&gt;&lt;tr&gt;&lt;td&gt;AlexNet&lt;/td&gt;&lt;td&gt;매우 간단&lt;/td&gt;&lt;td&gt;높음&lt;/td&gt;&lt;td&gt;적음&lt;/td&gt;&lt;td&gt;낮음&lt;/td&gt;&lt;td align=&quot;center&quot;&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;VGGNet&lt;/td&gt;&lt;td&gt;간단&lt;/td&gt;&lt;td&gt;높음&lt;/td&gt;&lt;td&gt;매우 높음&lt;/td&gt;&lt;td&gt;높음&lt;/td&gt;&lt;td align=&quot;center&quot;&gt;사용 용이&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;GoogLeNet&lt;/td&gt;&lt;td&gt;복잡&lt;/td&gt;&lt;td&gt;낮음&lt;/td&gt;&lt;td&gt;보통&lt;/td&gt;&lt;td&gt;매우 높음&lt;/td&gt;&lt;td align=&quot;center&quot;&gt;사용하기 어려움&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;ResNet&lt;/td&gt;&lt;td&gt;보통&lt;/td&gt;&lt;td&gt;보통&lt;/td&gt;&lt;td&gt;보통&lt;/td&gt;&lt;td&gt;높음&lt;/td&gt;&lt;td align=&quot;center&quot;&gt;사용 용이&lt;/td&gt;&lt;/tr&gt;&lt;/tbody&gt;&lt;/table&gt;&lt;hr/&gt;&lt;h3 id=&quot;reference&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#reference&quot; aria-label=&quot;reference permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Reference&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://www.javatpoint.com/k-nearest-neighbor-algorithm-for-machine-learning&quot;&gt;K-Nearest Neighbor(KNN) Algorithm for Machine Learning - Javatpoint&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[Computer Vision 02 - 효율적인 데이터 학습을 위한 Annotation]]></title><description><![CDATA[Annotation data efficient learning by 오태현 교수님, BoostCamp AI Tech 7주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/38_annotation_data_efficient_learning/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/38_annotation_data_efficient_learning/</guid><pubDate>Mon, 08 Mar 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;annotation-data-efficient-learning&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#annotation-data-efficient-learning&quot; aria-label=&quot;annotation data efficient learning permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Annotation data efficient learning&lt;/h1&gt;&lt;p&gt;실전에서 딥러닝을 사용할 때 가장 어려운 문제는 데이터의 레이블링이다. 레이블이 잘 붙여진 대량의 고품질 데이터가 있으면 효율적인 supervised learning을 수행할 수 있겠지만, 현실에서는 그런 경우가 거의 없거나 이를 구축하는 데에 비용이 매우 많이 든다.&lt;/p&gt;&lt;p&gt;따라서 주어진 데이터셋만으로 실제 데이터 분포를 따라가거나, 기학습된 정보를 이용하여 새로운 데이터셋을 더 효율적으로 학습하거나, 레이블이 없는 데이터셋까지 학습하는 다양한 data annotation 기법들을 살펴보자.&lt;/p&gt;&lt;h2 id=&quot;data-augmentation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#data-augmentation&quot; aria-label=&quot;data augmentation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Data Augmentation&lt;/h2&gt;&lt;p&gt;데이터셋은 항상 편향되어있다. 이미지 데이터를 예로 들자면, 구도, 사진사의 영향 등에 영향을 받기 때문에, 인간이 보기에 좋은 방향으로 치우쳐져 있다. 또 데이터셋은, 현실 데이터셋의 일부분(fractal)에 불과하므로 너무 sparse하기도 하다. 이처럼 학습에 사용되는 샘플 데이터셋과 현실 데이터셋 사이의 격차를 어떻게 메울 수 있을까?&lt;/p&gt;&lt;h3 id=&quot;image-data-augmentation-methodstechniques&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#image-data-augmentation-methodstechniques&quot; aria-label=&quot;image data augmentation methodstechniques permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Image Data Augmentation methods/techniques&lt;/h3&gt;&lt;p&gt;잘라내기(Crop), 비틀기(Shear), 밝기 조절(Brightness),  원근 변환(Perspective), 회전(Rotate), 반전(Flip), 기하학 변환 등의 다양한 이미지 변환 기법이 사용된다.&lt;/p&gt;&lt;p&gt;OpenCV와 NumPy 라이브러리가 이러한 방식들을 지원하고 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;밝기조절(Brightness)&lt;/code&gt;&lt;ul&gt;&lt;li&gt;각 픽셀별 RGB 값에 동일한 값을 더하거나, 적당한 값을 곱하는 등으로 scaling해준다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;회전/반전(Rotate/Flip)&lt;/code&gt;&lt;ul&gt;&lt;li&gt;라이브러리에서 rotate 기능을 지원한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;잘라내기(Crop)&lt;/code&gt;&lt;ul&gt;&lt;li&gt;사진에서 중요한 파트에 대해 더 강하게 학습할 수 있게 만들어주는 역할을 한다.&lt;/li&gt;&lt;li&gt;픽셀 범위 인덱싱으로 수행한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;기하학 변환&lt;ul&gt;&lt;li&gt;&lt;code&gt;Affine transformation&lt;/code&gt;&lt;ul&gt;&lt;li&gt;이미지의 선(line), 길이 비율(length ratio), 평행성(parallelism)은 보존한다.&lt;/li&gt;&lt;li&gt;이미지의 네 꼭짓점 중 세 꼭짓점을 mapping 대응쌍을 AffineTransform 함수에 넣어준다.&lt;/li&gt;&lt;li&gt;비틀거나, 회전하거나, 옮겨주는 기하학적 방법을 warping이라고 한다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;Shear&lt;/code&gt; transform이라고도 부른다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;최근의 변환 기법들&lt;ul&gt;&lt;li&gt;&lt;code&gt;CutMix&lt;/code&gt;&lt;ul&gt;&lt;li&gt;Cut과 Mix를 모두 사용하여, 잘라낸 각각의 두 사진을 이어붙여 학습데이터로 사용한다.. 이 때, 레이블도 동일한 비율로 조정한다.&lt;/li&gt;&lt;li&gt;이를 통해 서로 다른 두 물체의 위치를 좀 더 정교하게 catch할 수 있게 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;RandAugmentation&lt;/code&gt;&lt;ul&gt;&lt;li&gt;여러가지 가능한 영상처리 기법을 어떤 방식으로 조합할 지 탐색하는 것. 즉, best sequence of augmentation을 자동으로 탐색하는 기술&lt;/li&gt;&lt;li&gt;파라미터 : [사용할 기법의 종류(which), 기법의 강도(magnitude) ]&lt;ul&gt;&lt;li&gt;N개의 기법들 사이에서 샘플링한 뒤 수행하여 성능을 비교한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;leveraging-pre-trained-information&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#leveraging-pre-trained-information&quot; aria-label=&quot;leveraging pre trained information permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Leveraging pre-trained information&lt;/h2&gt;&lt;p&gt;데이터를 적게 쓰고도 좋은 성능을 발휘하기 위해서, 다른 데이터 셋에서 학습된, 즉 pre-trained 정보를 활용 해볼 수 있다.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;transfer-learning&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#transfer-learning&quot; aria-label=&quot;transfer learning permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Transfer learning&lt;/h3&gt;&lt;p&gt;Supervised learning에 사용되는 대용량의 데이터셋은 매우 비싸기도 하고, 사람이 만들었을 때 그 퀄리티를 보장하기 어렵다. 이를 해결하기 위한 실용적인 방법의 필요성이 대두되었다.&lt;/p&gt;&lt;p&gt;예를 들어, 주어진 데이터셋에서 오토바이의 바퀴를 찾아내고 싶다고 하자. 그런데 오토바이의 바퀴는 자동차의 바퀴와 비교적 비슷하게 생겼다. 따라서, 자동차의 사진들을 학습하여 바퀴를 구별해 낼 수 있는 기학습된 모델이 있다면,  해당 모델의 기학습된 정보를 이용하는 것이 좋을 것이다. 이런 경우에 &lt;strong&gt;&lt;code&gt;전이학습(Transfer learning)&lt;/code&gt;&lt;/strong&gt;을 사용한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:30.859375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAABGUlEQVQY032QzUsCURTF5y8uC12EtCkEYaaNCwuG1kJEIk0t1LAocKEplYxECjOCoiXzZpih+bjvvtu80aBVZ/E25/wO71xFCOJA/4hzCsI0RYnjxZ9u4AdhHG8sBRKqVoWu43wusqhIYFsmCSJ7Jo6OsV7nzev13Y3T7w6s0YiAZzBQuYy5HFYq/PSMDGP88TbAKApDMk1aLCS8s4vFA7i6YPctt9N6fB+bm2IJqyrm86iqXDuhWm348txNvlNRr4fTqUjhvX08LELjkj20Waf9ZM3s328DaRqWSjiZyDLPC5arVZStQpQJyxaFAp7rcNtwmoYz7L/atoWZp6Sv523n/ZU8pNxFYUTMTW0RsMh34vWXlO/7APADygZCFNxpw+oAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;transfer-learning&quot; title=&quot;transfer-learning&quot; src=&quot;/static/902ab85bb51864dafa97bb7852f125b1/2bef9/transfer-learning.png&quot; srcSet=&quot;/static/902ab85bb51864dafa97bb7852f125b1/6f3f2/transfer-learning.png 256w,/static/902ab85bb51864dafa97bb7852f125b1/01e7c/transfer-learning.png 512w,/static/902ab85bb51864dafa97bb7852f125b1/2bef9/transfer-learning.png 1024w,/static/902ab85bb51864dafa97bb7852f125b1/71c1d/transfer-learning.png 1536w,/static/902ab85bb51864dafa97bb7852f125b1/a878e/transfer-learning.png 2048w,/static/902ab85bb51864dafa97bb7852f125b1/ae628/transfer-learning.png 2604w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;기학습된 모델에서 전연결 계층(Fully connected layer)과 분류출력 부분만 제거하고, 그 대신에 새로운 전연결 계층과 내가 원하는 형태로 분류해줄 출력층을 정의한다. 이렇게 마지막 파트만 바꾼 모델을 다시 학습시킨다. 이 때, 합성곱 계층부(Convolution layers)는 Weight를 고정한다. 그래야 기존에 학습된 모든 정보들을 다 가지고 있기 때문이다. &lt;strong&gt;Pre-trained task에서 new task로 knowledge를 transfer하는 과정&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:910px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:89.84375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;transfer-learning2&quot; title=&quot;transfer-learning2&quot; src=&quot;/static/95856649d458f8c0ec4099086c444add/c6bbc/transfer-learning2.png&quot; srcSet=&quot;/static/95856649d458f8c0ec4099086c444add/6f3f2/transfer-learning2.png 256w,/static/95856649d458f8c0ec4099086c444add/01e7c/transfer-learning2.png 512w,/static/95856649d458f8c0ec4099086c444add/c6bbc/transfer-learning2.png 910w&quot; sizes=&quot;(max-width: 910px) 100vw, 910px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;또는, 합성곱 계층부는 낮은 학습률로, 새로 바꾼 전연결 계층부는 높은 학습률로 설정하여 전체(whole model)를 모두 업데이트시키는 &lt;strong&gt;Fine-tuning 방법&lt;/strong&gt;도 있다. 이 경우 부분만 학습하는 방식보다 조금 더 많은 데이터가 필요하겠으나, 당연히 성능은 더 좋을 것이다.&lt;/p&gt;&lt;p&gt;이 두 가지의 고전적인 방법과 더불어, 최근에는 좀 더 진보된 방식이 사용되고 있다.&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;knowledge-distillation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#knowledge-distillation&quot; aria-label=&quot;knowledge distillation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Knowledge Distillation&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:23.046875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAFCAIAAADKYVtkAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA60lEQVQY002PMYsCMRCF82vt7vAqQfwHdoKF4FmIhaAiK3iVnQoiWIkLKqJbLQquFi6YTOaNk7NxCI98j5eZiZH/AoJeLjKfY7HAdIo0DZb3cr3K+Sx5Lp9JVT0mSbBe435Xh8d/rlDgSoVVRyOvb7dbKpW4+MPLJYk4a50IAfbdyJTL/P3Fp0SZDgf0ehgO0e1ivw8jsiy03mzweCjidkO/z1H0HAx4tYKZTDCbwXswe+d8FNFvMySYQzrLuN3mRoPjOGCeI471ToppKkY+igidDleraLVgbVh7t6NaDfU6jidWJAqmfvCdfwFZzwyq2MVsawAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;knowledge-distillation&quot; title=&quot;knowledge-distillation&quot; src=&quot;/static/f7aeacbd27dd547eabe8ed41d44f2860/2bef9/knowledge-distillation.png&quot; srcSet=&quot;/static/f7aeacbd27dd547eabe8ed41d44f2860/6f3f2/knowledge-distillation.png 256w,/static/f7aeacbd27dd547eabe8ed41d44f2860/01e7c/knowledge-distillation.png 512w,/static/f7aeacbd27dd547eabe8ed41d44f2860/2bef9/knowledge-distillation.png 1024w,/static/f7aeacbd27dd547eabe8ed41d44f2860/71c1d/knowledge-distillation.png 1536w,/static/f7aeacbd27dd547eabe8ed41d44f2860/a878e/knowledge-distillation.png 2048w,/static/f7aeacbd27dd547eabe8ed41d44f2860/f36fd/knowledge-distillation.png 2488w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Knowledge distillation&lt;/code&gt;&lt;/strong&gt;은 대형 모델이 학습한 정보를 더 작은 별개의(another) 소형 모델로 전달해주는 방법이며, &lt;code&gt;Teacher-student learning&lt;/code&gt;이라고도 불린다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이는 모델 압축(model compression)에도 유용한 방법이다. 소형 모델이 대형 모델의 방식을 따라하기(mimicking) 때문이다.&lt;/li&gt;&lt;li&gt;최근에는 teacher model의 출력 레이블(예측)을 ground truth 레이블인 것 처럼 소형 모델에게 학습시키는 &lt;code&gt;pseudo-labeling&lt;/code&gt; 방식으로도 사용되고 있다. 레이블이 없는 데이터에 가짜 레이블(pseudo-label)을 붙이기 때문에 붙여진 이름이다. 따라서 unsupervised learning의 일종으로도 볼 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:33.203125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;knowledge-distillation-with-labeld-data&quot; title=&quot;knowledge-distillation-with-labeld-data&quot; src=&quot;/static/07cb0c191ddd061eff633114d71ca014/2bef9/knowledge-distillation-with-labeld-data.png&quot; srcSet=&quot;/static/07cb0c191ddd061eff633114d71ca014/6f3f2/knowledge-distillation-with-labeld-data.png 256w,/static/07cb0c191ddd061eff633114d71ca014/01e7c/knowledge-distillation-with-labeld-data.png 512w,/static/07cb0c191ddd061eff633114d71ca014/2bef9/knowledge-distillation-with-labeld-data.png 1024w,/static/07cb0c191ddd061eff633114d71ca014/71c1d/knowledge-distillation-with-labeld-data.png 1536w,/static/07cb0c191ddd061eff633114d71ca014/a878e/knowledge-distillation-with-labeld-data.png 2048w,/static/07cb0c191ddd061eff633114d71ca014/8b95f/knowledge-distillation-with-labeld-data.png 2848w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;strong&gt;*위의 그림에서 Student Loss를 예측하는 부분은 Soft prediction이 아닌 Hard prediction으로, 오기입니다.&lt;/strong&gt;&lt;/p&gt;&lt;p&gt;만약 레이블링이 된 데이터가 어느정도 있다면 있다면, 위와 같은 방식으로 student model의 학습을 두 갈래로 나누어 실제(ground truth) 레이블 데이터의 학습 비중을 좀 더 높이는 전략을 취할수도 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;Distillation Loss&lt;/code&gt; &lt;strong&gt;:&lt;/strong&gt; pseudo-labeling을 이용한 가짜 레이블과의 Loss&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;em&gt;KL-Divergence(&lt;/em&gt;&lt;/strong&gt;Soft label, Soft prediction)&lt;/li&gt;&lt;li&gt;Teacher 모델 추론과 Student 모델 추론의 차이&lt;/li&gt;&lt;li&gt;즉, Teacher Network가 알고있는 knowledge를 (흉내냄으로써) 학습한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;Student Loss&lt;/code&gt; &lt;strong&gt;:&lt;/strong&gt; 실제 레이블 데이터와의 Loss&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;em&gt;CrossEntropy&lt;/em&gt;&lt;/strong&gt;(Hard label, Soft Predction)&lt;/li&gt;&lt;li&gt;Student Network의 추론과 실제 레이블간의 차이&lt;/li&gt;&lt;li&gt;정답(right answer)을 학습한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:531px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:46.09375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;hard-vs-soft-label&quot; title=&quot;hard-vs-soft-label&quot; src=&quot;/static/45fc1e14ffbdb0a78233d0682fd07a9d/d4713/hard-vs-soft-label.png&quot; srcSet=&quot;/static/45fc1e14ffbdb0a78233d0682fd07a9d/6f3f2/hard-vs-soft-label.png 256w,/static/45fc1e14ffbdb0a78233d0682fd07a9d/01e7c/hard-vs-soft-label.png 512w,/static/45fc1e14ffbdb0a78233d0682fd07a9d/d4713/hard-vs-soft-label.png 531w&quot; sizes=&quot;(max-width: 531px) 100vw, 531px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;두 가지 Loss를 구하는 경우 중, Distillation Loss를 구할 때에는, 학습에 있어서 hard label 대신 soft label을 사용하는 &lt;strong&gt;Soft Prediction&lt;/strong&gt;을 수행한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;em&gt;Hard label(one-hot vector)&lt;/em&gt; :, 해당 클래스가 &amp;#x27;진짜&amp;#x27; 정답인지 아닌지만을 명확히 판별하는 레이블. 일반적으로 데이터셋이 보유한 ground truth 레이블이다.&lt;/li&gt;&lt;li&gt;&lt;em&gt;Soft label&lt;/em&gt; : 해당 데이터에 대해 모델이 어떻게 생각하는지, 즉 모델의 knowledge를 나타내주는 레이블이다. 일반적으로 모델의 output, 즉 추론한 결과이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:25.390625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAFCAIAAADKYVtkAAAACXBIWXMAABYlAAAWJQFJUiTwAAAAnklEQVQY022Pyw6DIBRE/f8/IyGywLXdsMHwtkVtQB2hJU3bs4HMvTMD3b7vzjmtdQhhnmdcvPfHB9CNMb6ABe+DUtq5a6erGyklSEqpbduOfyBZSoloNDWxQzMOIQRjjBAyjrdf57qunHNKad/3wzDEGCHC+Go2hWmakF3jvsyqgHI0t9dd5mVZjLH3Aj5W/9PIOVtrH28wtRb+J0Yn4uofDX1vrZAAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;softmax-with-temperature&quot; title=&quot;softmax-with-temperature&quot; src=&quot;/static/6e0b64b88297c8f0c349facd3e418421/2bef9/softmax-with-temperature.png&quot; srcSet=&quot;/static/6e0b64b88297c8f0c349facd3e418421/6f3f2/softmax-with-temperature.png 256w,/static/6e0b64b88297c8f0c349facd3e418421/01e7c/softmax-with-temperature.png 512w,/static/6e0b64b88297c8f0c349facd3e418421/2bef9/softmax-with-temperature.png 1024w,/static/6e0b64b88297c8f0c349facd3e418421/71c1d/softmax-with-temperature.png 1536w,/static/6e0b64b88297c8f0c349facd3e418421/a878e/softmax-with-temperature.png 2048w,/static/6e0b64b88297c8f0c349facd3e418421/1cab1/softmax-with-temperature.png 2444w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;Soft Prediction을 위해서는 모델의 추론 값을 Softmax에 통과시켜야 하는데, 이 때 &lt;strong&gt;T&lt;/strong&gt;, 즉 &lt;strong&gt;Tempature&lt;/strong&gt;라는 새로운 개념이 등장한다. 단순히 softmax만을 취하면, 입력의 값을 극단적으로 0 또는 1에 가깝게 벌려주게 된다. 이 대신에, 입력값을 temperature라는 상수로 나누어준 뒤 softmax를 취하면 output value를 어느정도 smooth하게 만들 수 있다. 이렇게 하여 output을 어느정도 입력(teacher model의 출력)을 따라가게 만듦으로써, student model이 teacher model을 더 잘 따라하게 할 수 있다.&lt;/p&gt;&lt;p&gt;또한, 기존에 학습된 teacher 모델의 task는 student 모델의 task와 전혀 다르기 때문에, teacher model의 각각의feature, 즉 dimension 정보는 그다지 중요하지 않다. 즉 각각의 element(node)가 가지는 semantic한 의미가 중요하지는 않다. 그보다는 teacher 모델의 행동을 추상적으로 정의하고 흉내내는 것이 student 모델의 목적이다. (&lt;em&gt;Semantic information is not considered in distillation&lt;/em&gt;)&lt;/p&gt;&lt;p&gt;이처럼 두 갈래로 나누어 학습하는 경우, Distillation Loss와 Student Loss를 가중합(Weighted Sum)하여 Student 모델을 학습한다.&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;leveraging-unlabeled-dataset-for-training&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#leveraging-unlabeled-dataset-for-training&quot; aria-label=&quot;leveraging unlabeled dataset for training permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Leveraging unlabeled dataset for training&lt;/h2&gt;&lt;p&gt;추가적인 레이블링 없이 성능을 올리는 방법에 대해서 살펴보자.&lt;/p&gt;&lt;h3 id=&quot;semi-supervised-learning&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#semi-supervised-learning&quot; aria-label=&quot;semi supervised learning permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Semi-supervised learning&lt;/h3&gt;&lt;p&gt;레이블이 있는 데이터는 supervised로, 레이블이 없는 데이터는 unsupervised로 학습하는 방식을 가리킨다. 레이블이 없는 데이터는 labeled data로 기학습된 모델을 이용하여 레이블이 없는 데이터들을 Pseudo-labeling하고, 이를 학습시킨다.&lt;/p&gt;&lt;h3 id=&quot;self-training&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#self-training&quot; aria-label=&quot;self training permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Self-training&lt;/h3&gt;&lt;p&gt;위의 방법들을 잘 결합하여 새로운 지평을 연 최신 연구로, &lt;strong&gt;Self-training with noisy student&lt;/strong&gt;라고도 한다. 기존의 Efficient Net에 비하여 큰 폭으로 성능이 향상되었다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:37.109375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;self-training&quot; title=&quot;self-training&quot; src=&quot;/static/b0251031be84f192f0bc0c8d8747a2c6/2bef9/self-training.png&quot; srcSet=&quot;/static/b0251031be84f192f0bc0c8d8747a2c6/6f3f2/self-training.png 256w,/static/b0251031be84f192f0bc0c8d8747a2c6/01e7c/self-training.png 512w,/static/b0251031be84f192f0bc0c8d8747a2c6/2bef9/self-training.png 1024w,/static/b0251031be84f192f0bc0c8d8747a2c6/71c1d/self-training.png 1536w,/static/b0251031be84f192f0bc0c8d8747a2c6/a878e/self-training.png 2048w,/static/b0251031be84f192f0bc0c8d8747a2c6/8b95f/self-training.png 2848w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ol&gt;&lt;li&gt;Teacher 모델이 실제 레이블을 토대로 가짜 레이블(pseudo-label)을 생성한다.&lt;/li&gt;&lt;li&gt;생성한 가짜 레이블과 실제 레이블을 섞어서(e.g. RandAugment) Student 모델을 학습시킨다.&lt;/li&gt;&lt;li&gt;기존의 Teacher 모델을 새로 만들어진 Student 모델로 갈아치운다. 즉, 학생이 자라서 선생이 된다.&lt;/li&gt;&lt;li&gt;1-3을 2~4회 정도 반복한다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;기존의 Knowledge Distillation에서는 Teacher 모델이 Student 모델보다 대형의 모델이었다. 그러나 이 방식을 사용하면, &lt;strong&gt;&lt;div&gt;각 round가 끝날때마다 Student 모델이 점점 더 커지게 된다.&lt;/div&gt;&lt;/strong&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[Graph 07 - 그래프 신경망(Graph Neural Network)]]></title><description><![CDATA[그래프 신경망 by 신기정 교수님, BoostCamp AI Tech 5주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/36_gnn/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/36_gnn/</guid><pubDate>Fri, 26 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;그래프-신경망&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B7%B8%EB%9E%98%ED%94%84-%EC%8B%A0%EA%B2%BD%EB%A7%9D&quot; aria-label=&quot;그래프 신경망 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;그래프 신경망&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;그래프 신경망(Graph Neural Network)&lt;/code&gt;&lt;/strong&gt;은 출력으로 인코더를 얻는 &lt;code&gt;귀납식 노드 임베딩&lt;/code&gt; 방법의 대표적인 사례이다.&lt;/p&gt;&lt;h2 id=&quot;gnn의-구조&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gnn%EC%9D%98-%EA%B5%AC%EC%A1%B0&quot; aria-label=&quot;gnn의 구조 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GNN의 구조&lt;/h2&gt;&lt;p&gt;그래프(의 인접행렬)와 정점의 속성 정보를 입력으로 받는다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;인접행렬 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(|V|\times|V|)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 이진행렬이다.&lt;/li&gt;&lt;li&gt;각 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 속성(Attribute) 벡터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;m&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;차원 벡터이고, 이 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;m&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;은 속성의 수를 의미한다.&lt;ul&gt;&lt;li&gt;속성 예시&lt;ul&gt;&lt;li&gt;온라인 소셜 네트워크에서의 사용자 지역, 성별, 연령, 프로필 사진 등&lt;/li&gt;&lt;li&gt;논문 인용 그래프에서 논문에 사용된 키워드에 대한 원-핫 벡터&lt;/li&gt;&lt;li&gt;PageRank 등의 정점 중심성, 군집 계수(Clustering Coefficient)&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:45.3125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAJCAIAAAC9o5sfAAAACXBIWXMAABYlAAAWJQFJUiTwAAABR0lEQVQoz4WRWU8DIRSF+/9/TV0e1dbGaGJs2kSNxtZ2hs7C1IECswKXQchMXR6MN0Buwv1yDoeR/VGdMfbv+rrtOtf6fuQ2AGjQ9r9yDBeisxZvbzMce9gAfOxzQpksBFne91P9qBP4Ro0JZxfj8als1ez9LozRoFyV9ECw0aaiuQL1W830JpXSnfRXWlbByzU9sAGuWcTzwDXokCzQk2fcUEOzZBOEURTFVVkp0PPd83R6UxZic3n+/rbycKtglYjep9SqWaKubL0dniVoHaAkipO6bsDANg+CENWkeDxbbAIvNqolrNPCy76mdM+HZDqLdjbBFpxr7900UqWkFmVhG9CINqo9pq01pRSne854z/bvdSkd47OY2JPJ6moyA12l0YMQpYcZ4ynGjLHODr93jMovrbWU0h2EAPcZ2ULoxZzizH/tJ8X5BAwgXHprAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;gnn-embedding&quot; title=&quot;gnn-embedding&quot; src=&quot;/static/2404ef4145cbd7c49f801c9e24584883/2bef9/gnn-embedding.png&quot; srcSet=&quot;/static/2404ef4145cbd7c49f801c9e24584883/6f3f2/gnn-embedding.png 256w,/static/2404ef4145cbd7c49f801c9e24584883/01e7c/gnn-embedding.png 512w,/static/2404ef4145cbd7c49f801c9e24584883/2bef9/gnn-embedding.png 1024w,/static/2404ef4145cbd7c49f801c9e24584883/71c1d/gnn-embedding.png 1536w,/static/2404ef4145cbd7c49f801c9e24584883/defc9/gnn-embedding.png 2046w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;그래프 신경망은 이웃 정점들의 정보를 집계하는 것을 반복하여 임베딩을 얻는다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;위의 이미지에서, 정점 A의 임베딩을 얻기 위해 A.neighborhood의 정보를 다시 집계한다.&lt;/li&gt;&lt;li&gt;A.neighborhood.neighborhood에서, A가 다시 등장하는 것을 확인할 수 있다.&lt;/li&gt;&lt;li&gt;이처럼, 각 집계 단계를 &lt;strong&gt;&lt;code&gt;층(Layer)&lt;/code&gt;&lt;/strong&gt; 이라 부르고, 각 층마다 임베딩을 얻는다. 각 층에서는 이웃들의 이전 층 임베딩을 집계하여 새로운 임베딩을 얻는다.&lt;ul&gt;&lt;li&gt;0번 층, 즉 입력층의 임베딩으로는 정점의 속성 벡터를 사용한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 때, 그래프에서 각 정점마다 이웃들이 다르므로 대상 정점마다 집계되는 정보가 상이하다. 이처럼 대상 정점별로 집계되는 구조를 계산 그래프(Computation Graph)라고 한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:35.9375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAABB0lEQVQY022Q227CMBBE8/+/1de+FKRCoAQoJVEDcby215f1hW5IBS+sLGvOSLMau7q9mpx8SfaFr+FWygOrGZLFZMxk3BGunehPt1xSCEpSDJSJcop9vSL7XFoV9qJzcjSif4SVQgOaCyQi1DERHyopjRJCCM9wjF6NP6yado9asxiu8Pb5dVi9j23HGEPk9SwubaOhW37XpfxXr87CmpANWBqwSDd3brbr3WZ72NeenOhD9IQ+LBYfOfjt50p3YqrI4ZQzAJyPnQLFlnNOQy9Om8uy0fJXGpgbnq/isKsRxuG4HnatRXt/cynEE6eLGRGlBKUUgNIajTZz2OGkJUh2XfD2/m1/6AaTGHrPYYIAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;gnn-embedding2&quot; title=&quot;gnn-embedding2&quot; src=&quot;/static/43abea42668b499f588d66de92e1b27d/2bef9/gnn-embedding2.png&quot; srcSet=&quot;/static/43abea42668b499f588d66de92e1b27d/6f3f2/gnn-embedding2.png 256w,/static/43abea42668b499f588d66de92e1b27d/01e7c/gnn-embedding2.png 512w,/static/43abea42668b499f588d66de92e1b27d/2bef9/gnn-embedding2.png 1024w,/static/43abea42668b499f588d66de92e1b27d/71c1d/gnn-embedding2.png 1536w,/static/43abea42668b499f588d66de92e1b27d/a878e/gnn-embedding2.png 2048w,/static/43abea42668b499f588d66de92e1b27d/88745/gnn-embedding2.png 2148w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;각 정점별 집계 정보가 다를지라도, 층 별 집계 함수는 공유된다. 정점 A나 정점 B나 0→1층에 사용되는 집계함수, 1→2층에 사용되는 집계함수는 동일하다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;서로 다른 층의 집계함수는 각기 다른 함수이다.&lt;/li&gt;&lt;li&gt;각 정점마다 연결성이 다르므로, 같은 집계함수를 사용하지만 입력의 수가 다를 수 있다. 즉, 집계함수는 &lt;strong&gt;서로 다른 구조의 계산 그래프를 처리할 수 있어야 한다&lt;/strong&gt;.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;집계-함수&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A7%91%EA%B3%84-%ED%95%A8%EC%88%98&quot; aria-label=&quot;집계 함수 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;집계 함수&lt;/h3&gt;&lt;p&gt;집계 함수의 구조는 다음과 같다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;이웃들 정보의 평균을 계산한다.&lt;ul&gt;&lt;li&gt;n개의 입력을 평균을 냄으로써 동일한 하나의 차원으로 축소할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;신경망에 적용한다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;이를 수식으로 나타내면 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/msubsup&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mfrac&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mstyle&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;msub&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mstyle&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∀&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;{&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;}&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
h_v^0 &amp;amp;= x_v\\
h_v^k &amp;amp;= \sigma\Bigg(\textcolor{green}{W_k\sum_{u\in N(v)}\frac{h_u^{k-1}}{|N(v)|}}+\textcolor{blue}{B_kh_v^{k-1}}\Bigg), \forall k \in \{1,\dots,K\}\\
z_v &amp;amp;= h_v^K
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:6.641444000000002em;vertical-align:-3.070722000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.5707220000000004em&quot;&gt;&lt;span style=&quot;top:-6.456614em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.75em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.046614em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.75em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.3392779999999993em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.75em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.070722000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.5707220000000004em&quot;&gt;&lt;span style=&quot;top:-6.456614em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.75em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.046614em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.75em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:green&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:green&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:green&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em;color:green&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot; style=&quot;color:green&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot; style=&quot;color:green&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;color:green&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.5261079999999998em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em;color:green&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:green&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:green&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;color:green;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8491079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:green&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:green&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em;color:blue&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.05017em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:blue&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:blue&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:blue&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:blue&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∀&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;}&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.3392779999999993em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.75em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.070722000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 임베딩을 의미한다.&lt;/li&gt;&lt;li&gt;0번층(입력층)의 임베딩은 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 속성 벡터로 초기화한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번층(현재층)의 임베딩은 1부터 층의 수(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;K&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)까지의  모든 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대하여 다음을 수행한다.&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mfrac&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{green}{W_k\sum_{u\in N(v)}\frac{h_u^{k-1}}{|N(v)|}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.6791199999999997em;vertical-align:-0.52em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:green&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:green&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;color:green;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.22528999999999993em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:green&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em;color:green&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot; style=&quot;color:green&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot; style=&quot;color:green&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.47471em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1591199999999997em&quot;&gt;&lt;span style=&quot;top:-2.655em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em;color:green&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot; style=&quot;color:green&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot; style=&quot;color:green&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;color:green;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.5102em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9270285714285713em&quot;&gt;&lt;span style=&quot;top:-2.214em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.931em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:green&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:green&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.52em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 이전 층에서 계산한 이웃들의 임베딩 평균을 선형변환한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;msub&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{blue}{B_kh_v^{k-1}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0961079999999999em;vertical-align:-0.247em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em;color:blue&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.05017em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:blue&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8491079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:blue&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:blue&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:blue&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 이전 층에서의 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 임베딩을 선형변환한다.&lt;/li&gt;&lt;li&gt;두 선형변환한 값의 합을 비선형변환(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sigma&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;마지막 층에서의 임베딩은 곧 출력 임베딩(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z_v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)이다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_k, B_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.05017em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 GNN이 학습하는 파라미터이다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;gnn의-학습&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gnn%EC%9D%98-%ED%95%99%EC%8A%B5&quot; aria-label=&quot;gnn의 학습 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GNN의 학습&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;1 - 먼저 손실함수를 정한다. 정점간 거리를 &amp;quot;보존&amp;quot;하는 것이 목표가 될 수 있다.&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;변환식 노드 임베딩처럼 인접성/거리 등을 기반으로 유사도를 정의하여 손실함수를 정의할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;또는, &lt;strong&gt;&lt;div&gt;&lt;code&gt;후속 과제(Downstream Task)&lt;/code&gt;의 손실함수를 이용한 &lt;code&gt;End-to-End 학습&lt;/code&gt;도 가능&lt;/div&gt;&lt;/strong&gt;하다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:46.484375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;end-to-end&quot; title=&quot;end-to-end&quot; src=&quot;/static/8f34e5f6cb240b3534cf84397f87f60e/2bef9/end-to-end.png&quot; srcSet=&quot;/static/8f34e5f6cb240b3534cf84397f87f60e/6f3f2/end-to-end.png 256w,/static/8f34e5f6cb240b3534cf84397f87f60e/01e7c/end-to-end.png 512w,/static/8f34e5f6cb240b3534cf84397f87f60e/2bef9/end-to-end.png 1024w,/static/8f34e5f6cb240b3534cf84397f87f60e/71c1d/end-to-end.png 1536w,/static/8f34e5f6cb240b3534cf84397f87f60e/de766/end-to-end.png 1838w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;예를 들어, Node Classification이 최종 목표일 경우, 아래와 같은 과정으로 학습할 수 있다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;GNN을 이용하여 노드 임베딩을 수행하고, 임베딩 값을 얻는다.&lt;/li&gt;&lt;li&gt;분류기의 입력으로 임베딩 값을 사용한다.&lt;/li&gt;&lt;li&gt;각 노드의 유형을 분류한다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;이 때 Downstream Task인 Classifier의 손실함수, Cross Entropy를 전체 프로세스의 손실함수로 사용할 수 있다. 아래의 식은 이진분류를 위한  Cross Entropy 식이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mstyle&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mstyle&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathcal{L} = \sum_{v\in V}\textcolor{green}{y_v}(\log(\sigma(z_v^\top\textcolor{red}{\theta})) + (1-\textcolor{green}{y_v})\log(1-\sigma(z_v^\top\textcolor{red}{\theta}))&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.3717110000000003em;vertical-align:-1.321706em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.8556639999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.321706em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em;color:red&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1491079999999998em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em;color:red&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{green}{y_v}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 정점의 실제 유형으로, 0 또는 1의 값을 갖는다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{red}{\theta}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em;color:red&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : Classifier의 학습 변수이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이처럼 Classifier의 손실함수를 토대로 Classfier뿐만 아니라 GNN 과정(Node Representation Learning)까지 역전파하여 한번에 학습할 수 있다.&lt;/p&gt;&lt;p&gt;End-to-End 학습을 통한 분류는 [Reductive Node Embedding(변환적 노드 임베딩) - Classifier 학습]보다 대체로 정확도가 더 높다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;2 - 학습에 사용할 대상 노드를 결정하여 학습 데이터를 구성한다.&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;선택한 일부 대상 노드들에 대해서만 계산 그래프를 구성한다. 이렇게 해도 되는 이유는, 집계함수가 공유되기 때문이다. 일부 노드에 대해서만 training하고, 다른 노드에 대해서 적용시킬 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;strong&gt;3 - 오차역전파를 통하여 손실함수를 최소화한다.&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 과정을 통해 신경망의 파라미터들을 학습한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;strong&gt;4 - 학습된 신경망을 다른 노드들에 적용하여, 학습에 사용되지 않은 노드의 임베딩을 얻을 수 있다. 또는, 아예 다른 그래프에 적용하여 임베딩을 얻어볼수도 있다.&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;학습 이후에 추가된 노드의 임베딩 또한 얻을 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;그래프-신경망-변형&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B7%B8%EB%9E%98%ED%94%84-%EC%8B%A0%EA%B2%BD%EB%A7%9D-%EB%B3%80%ED%98%95&quot; aria-label=&quot;그래프 신경망 변형 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;그래프 신경망 변형&lt;/h2&gt;&lt;p&gt;지금까지 보았던 GNN은 가장 기초적인 형태였고, 최근에는 이를 변형한 다양한 방법들이 대세를 이루고 있다. 가장 자주 사용되는 두 형태, 그래프 합성곱 신경망과  GraphSAGE를 알아보자.&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;그래프-합성곱-신경망&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B7%B8%EB%9E%98%ED%94%84-%ED%95%A9%EC%84%B1%EA%B3%B1-%EC%8B%A0%EA%B2%BD%EB%A7%9D&quot; aria-label=&quot;그래프 합성곱 신경망 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;그래프 합성곱 신경망&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;그래프 합성곱 신경망(Graph Convolution Network, GCN)&lt;/code&gt;&lt;/strong&gt;의 집계함수는 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/msubsup&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∪&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mfrac&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;msqrt&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;/msqrt&gt;&lt;/mfrac&gt;&lt;/mstyle&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∀&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;{&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;}&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
h_v^0 &amp;amp;= x_v \\
h_v^k &amp;amp;= \sigma\Bigg(W_k\textcolor{red}{\sum_{u \in N(v)\cup v}\frac{h_u^{k-1}}{\sqrt{|N(u)||N(v)|}}}\Bigg), \forall k \in \{1,\dots,K\}\\
z_v &amp;amp;= h_v^K
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:6.641444000000002em;vertical-align:-3.070722000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.5707220000000004em&quot;&gt;&lt;span style=&quot;top:-6.456614em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.75em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.046614em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.75em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.3392779999999993em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.75em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.070722000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.5707220000000004em&quot;&gt;&lt;span style=&quot;top:-6.456614em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.75em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.046614em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.75em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:red&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em;color:red&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:red&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:red&quot;&gt;∪&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:red&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;color:red&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.5261079999999998em&quot;&gt;&lt;span style=&quot;top:-2.175em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord sqrt&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.935em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3.2em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.2em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red;padding-left:1em&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em;color:red&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em;color:red&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:red&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.8950000000000005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.2em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail&quot; style=&quot;color:red;min-width:1.02em;height:1.28em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;1.28em&quot; viewBox=&quot;0 0 400000 1296&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M263,681c0.7,0,18,39.7,52,119
c34,79.3,68.167,158.7,102.5,238c34.3,79.3,51.8,119.3,52.5,120
c340,-704.7,510.7,-1060.3,512,-1067
l0 -0
c4.7,-7.3,11,-11,19,-11
H40000v40H1012.3
s-271.3,567,-271.3,567c-38.7,80.7,-84,175,-136,283c-52,108,-89.167,185.3,-111.5,232
c-22.3,46.7,-33.8,70.3,-34.5,71c-4.7,4.7,-12.3,7,-23,7s-12,-1,-12,-1
s-109,-253,-109,-253c-72.7,-168,-109.3,-252,-110,-252c-10.7,8,-22,16.7,-34,26
c-22,17.3,-33.3,26,-34,26s-26,-26,-26,-26s76,-59,76,-59s76,-60,76,-60z
M1001 80h400000v40h-400000z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30499999999999994em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;color:red;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8491079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:red&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:red&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.13em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∀&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;}&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.3392779999999993em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.75em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.070722000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;일견 기존 GNN과 큰 변화가 없어 보이지만, 다음과 같은 차이점들이 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;B_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.05017em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 파라미터가 있던 이전 층에서의 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 임베딩 값을 삭제하고, 동일 신경망을 사용하며  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 파라미터 하나만 학습을 수행한다.&lt;/li&gt;&lt;li&gt;정규화 방법(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sum&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.00001em;vertical-align:-0.25001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 분모)이 기존에는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 연결성이었다면, 지금은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 기하평균을 사용하고 있다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;graphsage&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#graphsage&quot; aria-label=&quot;graphsage permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GraphSAGE&lt;/h3&gt;&lt;p&gt;GraphSAGE의 집계함수는 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msubsup&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;mtext&gt;AGG&lt;/mtext&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;mo stretchy=&quot;false&quot;&gt;{&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∀&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;}&lt;/mo&gt;&lt;/mstyle&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;msub&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mstyle&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_v^k = \sigma([W_k\cdot \text{AGG}(\textcolor{blue}{\{h_u^{k-1},\forall u \in N(v)\}}), \textcolor{green}{B_kh_v^{k-1}}])&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.146108em;vertical-align:-0.247em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1491079999999998em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;AGG&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:blue&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:blue&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:blue&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:blue&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;∀&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot; style=&quot;color:blue&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1491079999999998em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em;color:blue&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:blue&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:blue&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:blue&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:blue&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em;color:green&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.05017em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:green&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:green&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:green&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;이웃들(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N(v)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)의 모든 임베딩을 &lt;strong&gt;AGG 함수&lt;/strong&gt;로 합친 다음,  자신의 임베딩(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;msub&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{green}{B_kh_v^{k-1}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0961079999999999em;vertical-align:-0.247em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em;color:green&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.05017em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:green&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8491079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:green&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:green&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)과 연결(Concatenation)한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 때 AGG 함수는 여러가지가 사용될 수 있다.&lt;/p&gt;&lt;table&gt;&lt;thead&gt;&lt;tr&gt;&lt;th&gt;AGG 함수&lt;/th&gt;&lt;th&gt;수식&lt;/th&gt;&lt;th&gt;비고&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;&lt;tbody&gt;&lt;tr&gt;&lt;td&gt;Mean(평균)&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mtext&gt;AGG&lt;/mtext&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mfrac&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\text{AGG} = \sum_{u\in N(v)}\frac{h_u^{k-1}}{\|N(v)\|}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;AGG&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.6791199999999997em;vertical-align:-0.52em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.22528999999999993em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.47471em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1591199999999997em&quot;&gt;&lt;span style=&quot;top:-2.655em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∥&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.5102em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9270285714285713em&quot;&gt;&lt;span style=&quot;top:-2.214em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.931em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.52em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/td&gt;&lt;td&gt;&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;Pool(풀링)&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;{&lt;/mo&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∀&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;}&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;AGG = \\gamma(\{Qh_u^{k-1}, \forall u \in N(v)\})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0991079999999998em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8491079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∀&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;γ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\gamma&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05556em&quot;&gt;γ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 원소별 최댓값을 의미&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;LSTM&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∀&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;π&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;AGG = LSTM([h_u^{k-1},\forall u \in \pi(N(v))])&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0991079999999998em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8491079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∀&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;π&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/td&gt;&lt;td&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;π&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\pi&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;π&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 순서를 섞는다는 의미&lt;/td&gt;&lt;/tr&gt;&lt;/tbody&gt;&lt;/table&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;합성곱-신경망과의-비교&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%95%A9%EC%84%B1%EA%B3%B1-%EC%8B%A0%EA%B2%BD%EB%A7%9D%EA%B3%BC%EC%9D%98-%EB%B9%84%EA%B5%90&quot; aria-label=&quot;합성곱 신경망과의 비교 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;합성곱 신경망과의 비교&lt;/h2&gt;&lt;p&gt;그렇다면 GNN와 CNN은 어떤 유사점과 차이점이 있을까?&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;유사성&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9C%A0%EC%82%AC%EC%84%B1&quot; aria-label=&quot;유사성 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;유사성&lt;/h3&gt;&lt;p&gt;둘 모두 이웃의 정보를 집계하는 과정을 반복한다. 구체적으로, CNN은 이웃 픽셀의 정보를 집계한다.&lt;/p&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;차이점&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%B0%A8%EC%9D%B4%EC%A0%90&quot; aria-label=&quot;차이점 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;차이점&lt;/h3&gt;&lt;p&gt;CNN에서는 이웃(픽셀)의 수가 균일하지만, GNN에서는 노드별로 집계하는 이웃의 수가 다르다.&lt;/p&gt;&lt;div&gt;&lt;p&gt;그래프의 인접 행렬에 CNN을 적용하면 효과적일까? 답은 &amp;#x27;그렇지 않다&amp;#x27;이다. \n CNN이 주로 사용되는 이미지는 인접 픽셀이 유용한 정보를 담고 있을 가능성이 높다. 그러나, 그래프의 인접행렬에서의 인접 원소는 제한된 정보만을 가진다. 특히, 인접 행렬의 행과 열 순서는 임의로 결정되는 경우가 많으므로, 인접 원소의 정보 유용성이 떨어진다. \n 따라서, 반드시 CNN이 아니라 GNN을 적용하여야 한다.&lt;/p&gt;&lt;/div&gt;&lt;h2 id=&quot;한계&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%95%9C%EA%B3%84&quot; aria-label=&quot;한계 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;한계&lt;/h2&gt;&lt;p&gt;GNN에서는 이웃들의 정보를 동일한 가중치로 평균을 낸다. GCN 또한 역시 단순히 연결성을 고려한 가중치로 평균을 낸다.&lt;/p&gt;&lt;p&gt;그러나 실제 그래프에서 모든 이웃들이 동등한 역할을 하는 것은 아니다. 실제 그래프에서는 이웃 별로 미치는 영향이 다를 수 있기 때문이다. 예를 들어 페이스북에서, 단순히 팔로우하는 친구보다 더 &amp;#x27;친한 친구&amp;#x27;가 있을 수 있다. 이런 것들을 고려할 수는 없을까?&lt;/p&gt;&lt;h2 id=&quot;그래프-어텐션attention-신경망&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B7%B8%EB%9E%98%ED%94%84-%EC%96%B4%ED%85%90%EC%85%98attention-%EC%8B%A0%EA%B2%BD%EB%A7%9D&quot; aria-label=&quot;그래프 어텐션attention 신경망 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;그래프 어텐션(Attention) 신경망&lt;/h2&gt;&lt;h3 id=&quot;self-attention&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#self-attention&quot; aria-label=&quot;self attention permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Self-Attention&lt;/h3&gt;&lt;p&gt;Graph Attention Network(GAT)는 (이웃과의 관계) 가중치 자체도 학습한다. 이 과정에서 self-attention이 사용된다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:29.6875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA/0lEQVQY00VQS07DMBDNJTkGC07AHRAbjlGEuEEFG8SGUqdtvs2nVqLU2HFtJ/4wjkV5sq2Zp3lvPBNxzoUQzlkH11pKqZQSYm0NHGM9zxgTl4tbMM8zISTURJCgeNedKSRcTtt4z8fR/WEycxBsETphAjHuaZYXSz8XwT0Pg9T2O87yvAoaRMvbj4e7z8eb9f1r+w4M+SHKuK/NrixP4Y9erJSqm5ZPrmpwUWOhFEzQKfJcr1+at6fDCrGjkXNRNVTYFnf5EY+MG2O8GIbcH1LtjcCeJkkCW7h6h7fv+7wo9VLUD0OapTCIF2utoXmoAwRLCMLC/M6cJ6+8df/4BYRUVHYaJM2QAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;gat-self-attention&quot; title=&quot;gat-self-attention&quot; src=&quot;/static/f46540bc0de3def1ab34693011c17c53/2bef9/gat-self-attention.png&quot; srcSet=&quot;/static/f46540bc0de3def1ab34693011c17c53/6f3f2/gat-self-attention.png 256w,/static/f46540bc0de3def1ab34693011c17c53/01e7c/gat-self-attention.png 512w,/static/f46540bc0de3def1ab34693011c17c53/2bef9/gat-self-attention.png 1024w,/static/f46540bc0de3def1ab34693011c17c53/71c1d/gat-self-attention.png 1536w,/static/f46540bc0de3def1ab34693011c17c53/a878e/gat-self-attention.png 2048w,/static/f46540bc0de3def1ab34693011c17c53/b7a14/gat-self-attention.png 2864w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;각 층의 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로부터 이웃 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;j&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.85396em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로의 가중치 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;a_{ij}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 다음과 같이 계산한다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;&lt;p&gt;해당 층의 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 임베딩 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 신경망 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{red}{W}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 곱해 새로운 임베딩을 얻는다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tilde{h}_i = h_i\textcolor{red}{W}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0812999999999997em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9312999999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.61344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;정점 i와 j의 새로운 임베딩을 연결한 후, 어텐션 계수 벡터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;purple&quot;&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{purple}{a}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:purple&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 를 내적한다. 어텐션 계수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;purple&quot;&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{purple}{a}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:purple&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 모든 정점이 공유하는 학습변수이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msup&gt;&lt;mstyle mathcolor=&quot;purple&quot;&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mstyle&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mtext&gt;CONCAT&lt;/mtext&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;e_{ij} = \textcolor{purple}{a}^\top[\text{CONCAT}(\tilde{h}_i,\tilde{h}_j)] &lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.2174079999999998em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:purple&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;CONCAT&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9312999999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.61344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9312999999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.61344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;e_{ij}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 softmax를 적용한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;exp&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi&gt;exp&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;a_{ij} = softmax_j(e_{ij}) = \frac{\exp(e_{ij})}{\sum_{k \in N_i}\exp(e_{ik})}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.51281em;vertical-align:-1.08581em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.18639799999999995em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.10903em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.39981em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;exp&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop&quot;&gt;exp&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.08581em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;이 과정에서 사용되는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mstyle&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mstyle mathcolor=&quot;purple&quot;&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{red}W, \textcolor{purple} a&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:purple&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 모두 오류역전파를 통한 학습의 결과로 얻어지는 파라미터들이다.&lt;/p&gt;&lt;h3 id=&quot;multi-head-attention&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#multi-head-attention&quot; aria-label=&quot;multi head attention permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Multi-head attention&lt;/h3&gt;&lt;p&gt;&lt;code&gt;멀티헤드 어텐션(Multi-head Attention&lt;/code&gt;)으로, 여러 개의 어텐션을 동시에 학습한 뒤, 결과를 연결하여 사용할수도 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo mathvariant=&quot;normal&quot; lspace=&quot;0em&quot; rspace=&quot;0em&quot;&gt;′&lt;/mo&gt;&lt;/msubsup&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mtext&gt;CONCAT&lt;/mtext&gt;&lt;/mo&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;≤&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;≤&lt;/mo&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;msubsup&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h&amp;#x27;_i = \underset{1\le k\le K}{\text{CONCAT}}\sigma\Bigg(\sum_{j \in N_i}a^k_{ij}h_jW_k\Bigg)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.048892em;vertical-align:-0.247em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8018919999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;′&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.180444em;vertical-align:-1.430444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.68333em&quot;&gt;&lt;span style=&quot;top:-2.347892em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;≤&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;≤&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;CONCAT&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8472869999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.8556639999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.10903em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.430444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.899108em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.1130000000000004em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.383108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;가중치 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;a_{ij}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개 얻어 집계한 결과를 concat하여 최종 임베딩값으로 활용한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;GAT를 적용한 결과, GCN보다도 유의미한 성능의 향상을 볼 수 있었다.&lt;/p&gt;&lt;h2 id=&quot;그래프-임베딩&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B7%B8%EB%9E%98%ED%94%84-%EC%9E%84%EB%B2%A0%EB%94%A9&quot; aria-label=&quot;그래프 임베딩 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;그래프 임베딩&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;그래프 표현 학습&lt;/code&gt;&lt;/strong&gt;, 혹은 &lt;strong&gt;&lt;code&gt;그래프 임베딩(Graph Embedding)&lt;/code&gt;&lt;/strong&gt;이란 그래프 전체를 벡터의 형태로 표현하는 것이다.&lt;/p&gt;&lt;p&gt;이는 개별 정점을 각각의 벡터로 표현하는 노드 임베딩과는 다른 개념이며, 그래프 임베딩은 벡터의 형태로 표현된 &lt;strong&gt;&amp;#x27;그래프 자체&amp;#x27;를 의미&lt;/strong&gt;하기도 한다.&lt;/p&gt;&lt;p&gt;그래프 임베딩은 Graph Classification등에 이용되며,  그래프 형태로 표현된 화합물의 분자구조로부터 특성 예측 등에 사용될 수 있다.&lt;/p&gt;&lt;h3 id=&quot;그래프-풀링&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B7%B8%EB%9E%98%ED%94%84-%ED%92%80%EB%A7%81&quot; aria-label=&quot;그래프 풀링 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;그래프 풀링&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;그래프 풀링(Graph Pooling)&lt;/code&gt;&lt;/strong&gt;이란 &lt;strong&gt;노드 임베딩들로부터 그래프 임베딩을 얻는 과정&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;평균 등의 단순한 방법보다, 그래프의 구조를 고려한 방법을 사용할 경우 그래프 Classification 등의 downstream task에서 더 높은 성능을 얻는다고 알려져 있다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;미분가능한 풀링(Differentiable Pooling, DiffPool)&lt;/code&gt;&lt;/strong&gt;은 군집 구조를 활용하여 임베딩을 계층적으로 집계한다.&lt;/p&gt;&lt;h2 id=&quot;over-smoothing-문제&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#over-smoothing-%EB%AC%B8%EC%A0%9C&quot; aria-label=&quot;over smoothing 문제 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Over-smoothing 문제&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;지나친 획일화(Over-smoothing) 문제&lt;/code&gt;&lt;/strong&gt;란 그래프 신경망 층의 수가 증가함에 따라 정점의 임베딩이 서로 유사해지는 현상을 말한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이는 small-world network와도 관련있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:21.09375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAECAIAAAABPYjBAAAACXBIWXMAABYlAAAWJQFJUiTwAAAAjElEQVQI1yVOSRLDMAzK/x/bpJW1WYuLHUYHJBjElZkRUVURE6hKDC7YIFV1ZLg7+ForDmBQGSAXhO7sCmFW1bXKjYX1mBzczM20toHmNMRVTdMvkq/9zs1dkaIipkI/8gPVQb+bWd5qboIseDOyu/dnlIGPiCDjNAbz4D7g3cXetvf9gfTy53lEBIY/wpbp8/WIm7IAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;over-smoothing&quot; title=&quot;over-smoothing&quot; src=&quot;/static/491f4850b016677edc9e308bba29a5f1/2bef9/over-smoothing.png&quot; srcSet=&quot;/static/491f4850b016677edc9e308bba29a5f1/6f3f2/over-smoothing.png 256w,/static/491f4850b016677edc9e308bba29a5f1/01e7c/over-smoothing.png 512w,/static/491f4850b016677edc9e308bba29a5f1/2bef9/over-smoothing.png 1024w,/static/491f4850b016677edc9e308bba29a5f1/71c1d/over-smoothing.png 1536w,/static/491f4850b016677edc9e308bba29a5f1/a878e/over-smoothing.png 2048w,/static/491f4850b016677edc9e308bba29a5f1/957b0/over-smoothing.png 3580w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위의 사진을 보면, 레이어를 층층이 쌓을수록 점점 노드들이 모이는 것을 볼 수 있다. 이는, 층을 쌓을 때마다 해당 노드를 임베딩하기 위해 확인하는 이웃 노드들의 범위가 커지기 때문이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;1층일 때는 해당 노드의 이웃만 확인한다.&lt;/li&gt;&lt;li&gt;2층일 때는 해당 노드의 이웃의 이웃도 확인한다.&lt;/li&gt;&lt;li&gt;3층일 때는 해당 노드의 이웃의 이웃의 이웃도 확인한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 과정이 반복되면서, 한 노드로 모이는 임베딩값은 결국 그래프의 넓은 부분을 커버하게 된다. 이에 따라 각 노드들의 특성 차이는 점점 희미해지고, 모든 노드들이 그래프의 대부분을 비슷하게 커버하기 때문에 비슷한 노드로 판별된다.&lt;/p&gt;&lt;p&gt;이처럼 그래프 신경망의 층의 수를 늘렸을 때, over-smoothing의 결과로 downstream task에서 정확도가 감소하는 현상이 발견되었다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;데이터에 따라 다르지만, 일반적으로 층을 2개 혹은 3개 정도 쌓을 때가 정확도가 가장 높다.&lt;/li&gt;&lt;li&gt;잔차항(Residual)을 넣어주는 것, 즉 이전 층의 임베딩을 한번 더 더해주는 것만으로는 효과가 제한적이다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;대안&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%8C%80%EC%95%88&quot; aria-label=&quot;대안 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;대안&lt;/h3&gt;&lt;p&gt;이러한 Over-smoothing problem의 대안으로 &lt;strong&gt;&lt;code&gt;JK 네트워크(Jumping Knowledge Network)&lt;/code&gt;&lt;/strong&gt;는 마지막 층의 임베딩 뿐 아니라, 모든 층의 임베딩을 최종 집계함수에 같이 넣어주어 최종 임베딩을 얻는 방식을 사용한다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;APPNP&lt;/code&gt;&lt;/strong&gt;는 0번째 층을 제외하고는 신경망(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;) 없이 집계함수를 단순화하였다. &lt;/p&gt;&lt;p&gt;두 방법 모두 효과가 있는 것으로 나타났으며, 특히 APPNP의 경우 node classification downstream task에서 층의 수 증가에 따른 정확도 감소 효과가 없는 것이 확인되었다.&lt;/p&gt;&lt;h2 id=&quot;그래프-데이터의-증강&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B7%B8%EB%9E%98%ED%94%84-%EB%8D%B0%EC%9D%B4%ED%84%B0%EC%9D%98-%EC%A6%9D%EA%B0%95&quot; aria-label=&quot;그래프 데이터의 증강 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;그래프 데이터의 증강&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;데이터 증강(Data Augmentation)&lt;/code&gt;&lt;/strong&gt;은 다양한 기계학습 문제에서 효과적이다. 특히 이미지 관련 문제에는 더더욱 효과적이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:30.078125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA3UlEQVQY02VPSXKEMAzk/7d8KF/gFB8g1DCZhXEVM8aLvMgY0nAISaVLsrtkt6Su1h3GUmIGIR9DjCApJR8CyIJYVhd4wfUXFdJapyYzz7Nz7qUm5hxCUEpBP+eMD2epe6mfLsUY8YT6IWZmIoIYxFqLM+eMRkTIIPoR4bm8vZ/Hp+pPJ/L+EGttjXXbzkQvpUsp6L1XFs7lOpIN23yeC2+APz7EMaam/RyG4etyeTzk3k7Xdd00Dda5327iQ9yHYf2HynuvdmAaLOXdpJSybVshhDFG66nruh+fv/ENJjdZJpBCzuIAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;data-augmentation&quot; title=&quot;data-augmentation&quot; src=&quot;/static/8fd20acb104a85337fac71f7be3aa53c/2bef9/data-augmentation.png&quot; srcSet=&quot;/static/8fd20acb104a85337fac71f7be3aa53c/6f3f2/data-augmentation.png 256w,/static/8fd20acb104a85337fac71f7be3aa53c/01e7c/data-augmentation.png 512w,/static/8fd20acb104a85337fac71f7be3aa53c/2bef9/data-augmentation.png 1024w,/static/8fd20acb104a85337fac71f7be3aa53c/71c1d/data-augmentation.png 1536w,/static/8fd20acb104a85337fac71f7be3aa53c/a878e/data-augmentation.png 2048w,/static/8fd20acb104a85337fac71f7be3aa53c/109e2/data-augmentation.png 3352w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;그래프에도 누락되거나 부정확한 간선이 있을 수 있는데, 이를 데이터 증강을 통해 보완할 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Random Walk를 통하여 정점간의 유사도를 계산한 뒤, 유사도가 높은 정점간의 간선을 연결하여 간선을 추가한 그래프를 GNN에 집어넣으면 더 효과가 좋다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;그래프에서 사용되는 데이터증강 기법으로는 HEAT, PPR 등이 있는데, 이러한 데이터 증강의 결과로 Node Classification의 정확도가 개선되는 것이 확인되었다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[Graph 05 - 노드 임베딩]]></title><description><![CDATA[그래프를 추천시스템에 어떻게 활용할까 by 신기정 교수님, BoostCamp AI Tech 5주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/34_node_embedding/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/34_node_embedding/</guid><pubDate>Thu, 25 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;노드-임베딩정점-임베딩&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%85%B8%EB%93%9C-%EC%9E%84%EB%B2%A0%EB%94%A9%EC%A0%95%EC%A0%90-%EC%9E%84%EB%B2%A0%EB%94%A9&quot; aria-label=&quot;노드 임베딩정점 임베딩 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;노드 임베딩(정점 임베딩)&lt;/h1&gt;&lt;h2 id=&quot;정점-표현-학습&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A0%95%EC%A0%90-%ED%91%9C%ED%98%84-%ED%95%99%EC%8A%B5&quot; aria-label=&quot;정점 표현 학습 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;정점 표현 학습&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;정점 표현 학습(Node Representation Learning)&lt;/code&gt;&lt;/strong&gt;은 &lt;strong&gt;그래프의 정점들을 벡터 형태로 표현하는 것&lt;/strong&gt;이다. 간단히 &lt;strong&gt;&lt;code&gt;노드 임베딩(Node Embedding)&lt;/code&gt;&lt;/strong&gt;이라고도 부른다.&lt;/p&gt;&lt;p&gt;노드 임베딩은 &lt;strong&gt;벡터 형태의 표현 그 자체&lt;/strong&gt;를 의미하기도 한다. 정점이 표현되는 벡터 공간을 &lt;code&gt;임베딩 공간&lt;/code&gt;이라고 하자.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;정점 표현 학습의 입력은 그래프이다.&lt;/li&gt;&lt;li&gt;정점 표현 학습의 출력은 벡터표현 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z_u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로, 각 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대한 임베딩을 의미한다.(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;R&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/msub&gt;&lt;mtext&gt;차원&lt;/mtext&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathbb{R}_d 차원&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83889em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;차&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;원&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;왜-노드-임베딩을-사용하는가&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%99%9C-%EB%85%B8%EB%93%9C-%EC%9E%84%EB%B2%A0%EB%94%A9%EC%9D%84-%EC%82%AC%EC%9A%A9%ED%95%98%EB%8A%94%EA%B0%80&quot; aria-label=&quot;왜 노드 임베딩을 사용하는가 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;왜 노드 임베딩을 사용하는가?&lt;/h3&gt;&lt;p&gt;노드 임베딩을 하면 그래프의 노드들을 벡터형태로 표현할 수 있게 된다. 즉, 벡터 형태의 데이터들을 위한 도구들을 그래프에도 적용할 수 있다. 그래프를 위한 별도의 알고리즘이 아니라, 기계학습의 방식을 그래프에서 사용할 수 있는 것이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;기계학습 도구들&lt;ul&gt;&lt;li&gt;Classifier(로지스틱 회귀분석, MLP 등)&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;최신 ML 기법들을 정점 분류(Node Classificaiton), 군집 분석(Community Detection) 등에 사용할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;code&gt;K-Means&lt;/code&gt;, &lt;code&gt;DBSCAN&lt;/code&gt; 같은 &lt;strong&gt;&lt;code&gt;군집 분석 알고리즘&lt;/code&gt;&lt;/strong&gt;들은 &lt;strong&gt;벡터형태로 표현된 사례(instance)&lt;/strong&gt;들을 입력으로 받는다.&lt;/p&gt;&lt;h3 id=&quot;노드임베딩의-목표&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%85%B8%EB%93%9C%EC%9E%84%EB%B2%A0%EB%94%A9%EC%9D%98-%EB%AA%A9%ED%91%9C&quot; aria-label=&quot;노드임베딩의 목표 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;노드임베딩의 목표&lt;/h3&gt;&lt;p&gt;노드 임베딩은 그래프에서의 정점간 유사도를 임베딩 공간에서도 보존할 수 있도록 수행되어야한다. 이 때, 유사도로는 내적(Inner Product)를 사용한다.&lt;/p&gt;&lt;p&gt;임베딩 공간에서의 노드 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 노드 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 &lt;code&gt;유사도&lt;/code&gt;는 &lt;strong&gt;두 노드의 임베딩 출력값의 내적값&lt;/strong&gt;과 같다. 내적은 두 벡터가 클 수록, 그리고 같은 방향을 향할 수록 큰 값을 갖는다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mtext&gt;similarity&lt;/mtext&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;≈&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;mi&gt;cos&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textrm{similarity}(u,v)\approx z^{\top}_vz_u = \Vert z_u\Vert\cdot\Vert z_v\Vert\cdot\cos(\theta)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;similarity&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.146108em;vertical-align:-0.247em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;cos&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;h2 id=&quot;인접성-기반-유사도&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9D%B8%EC%A0%91%EC%84%B1-%EA%B8%B0%EB%B0%98-%EC%9C%A0%EC%82%AC%EB%8F%84&quot; aria-label=&quot;인접성 기반 유사도 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;인접성 기반 유사도&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;인접성 기반 유사도(Adjacency-based similarity)&lt;/code&gt;&lt;/strong&gt;에서는 &lt;strong&gt;두 정점이 인접할 때 유사하다고 간주&lt;/strong&gt;한다. 두 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 인접하다는 것은 둘을 직접 연결하는 간선 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(u,v)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 있음을 의미한다.&lt;/p&gt;&lt;p&gt;이를 인접행렬(Adjacency Matrix) &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 표현하였을 때, 두 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 사이의 간선 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(u,v)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A_{u,v}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.969438em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 나타낼 수 있으며, 인접하였을 경우 이 값이 1이고 그렇지 않으면 0이다. 따라서, 인접행렬의 원소 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A_{u,v}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.969438em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 &lt;strong&gt;두 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 유사도&lt;/strong&gt;로 나타낼 수 있다.&lt;/p&gt;&lt;h3 id=&quot;손실-함수&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%86%90%EC%8B%A4-%ED%95%A8%EC%88%98&quot; aria-label=&quot;손실 함수 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;손실 함수&lt;/h3&gt;&lt;p&gt;인접성 기반 접근법의 손실 함수(Loss Function)는 다음과 같으며, 손실을 최소화하는 노드 임베딩을 찾기 위해 SGD등이 사용된다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mstyle mathcolor=&quot;purple&quot;&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msub&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathcal{L} = \sum_{(u,v)\in V\times V}\Vert \textcolor{purple}{z_u^{\top}z_v}-\textcolor{red}{A_{u,v}}\Vert^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.5660100000000003em;vertical-align:-1.516005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em;color:purple&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:purple&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em;color:purple&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.150216em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot; style=&quot;color:red&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:red&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;purple&quot;&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{purple}{z_u^{\top}z_v}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.096108em;vertical-align:-0.247em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em;color:purple&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.849108em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:purple&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em;color:purple&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 임베딩 공간에서의 유사도&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msub&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{red}{A_{u,v}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.969438em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot; style=&quot;color:red&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:red&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 그래프에서의 유사도&lt;/li&gt;&lt;li&gt;임베딩 공간의 유사도와 그래프에서의 유사도 차이가 최소가 되도록 한다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;한계&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%95%9C%EA%B3%84&quot; aria-label=&quot;한계 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;한계&lt;/h3&gt;&lt;p&gt;그러나 인접성 기반의 접근방법은 거리가 1보다 큰 정점들의 유사도를 0으로밖에 계산하지 못하므로, 정점들간의 유사도를 제대로 비교하기에는 무리가 한계가 있다. 또한, 어느 군집에 속했는가(같은 군집인지 아닌지 여부)도 유사도에 반영되지 않는다.&lt;/p&gt;&lt;h2 id=&quot;거리경로중첩-기반-유사도&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B1%B0%EB%A6%AC%EA%B2%BD%EB%A1%9C%EC%A4%91%EC%B2%A9-%EA%B8%B0%EB%B0%98-%EC%9C%A0%EC%82%AC%EB%8F%84&quot; aria-label=&quot;거리경로중첩 기반 유사도 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;거리/경로/중첩 기반 유사도&lt;/h2&gt;&lt;h3 id=&quot;거리-기반-유사도&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B1%B0%EB%A6%AC-%EA%B8%B0%EB%B0%98-%EC%9C%A0%EC%82%AC%EB%8F%84&quot; aria-label=&quot;거리 기반 유사도 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;거리 기반 유사도&lt;/h3&gt;&lt;p&gt;&lt;code&gt;거리 기반 유사도(Distance-based Similarity)&lt;/code&gt;에서는 &lt;strong&gt;두 정점 사이의 거리가 &amp;#x27;충분히&amp;#x27; 가까운 경우 유사&lt;/strong&gt;하다고 간주한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때, &amp;#x27;충분히&amp;#x27;의 기준은 임의적이다(하이퍼파라미터)&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;경로-기반-유사도&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B2%BD%EB%A1%9C-%EA%B8%B0%EB%B0%98-%EC%9C%A0%EC%82%AC%EB%8F%84&quot; aria-label=&quot;경로 기반 유사도 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;경로 기반 유사도&lt;/h3&gt;&lt;p&gt;&lt;code&gt;경로 기반 유사도(Path-based Similarity)&lt;/code&gt;에서는 &lt;strong&gt;두 정점 사이의 경로가 많을수록 유사&lt;/strong&gt;하다고 간주한다.&lt;/p&gt;&lt;p&gt;두 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 사이의 경로 중 거리가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;인 것의 수는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A_{u,v}^k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.232216em;vertical-align:-0.383108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.849108em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.383108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, 즉 인접행렬 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;제곱 행렬의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;행 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;열 원소와 같다.&lt;/p&gt;&lt;p&gt;경로 기반 접근법의 손실함수는 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mstyle&gt;&lt;/msubsup&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathcal{L} = \sum_{(u,v)\in V\times V}\Vert z_u^{\top}z_v-A_{u,v}^\textcolor{red}k\Vert^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.5660100000000003em;vertical-align:-1.516005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.282216em;vertical-align:-0.383108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.899108em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.1130000000000004em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em;color:red&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.383108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;h3 id=&quot;중첩-기반-유사도&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A4%91%EC%B2%A9-%EA%B8%B0%EB%B0%98-%EC%9C%A0%EC%82%AC%EB%8F%84&quot; aria-label=&quot;중첩 기반 유사도 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;중첩 기반 유사도&lt;/h3&gt;&lt;p&gt;&lt;code&gt;중첩 기반 유사도(Neighborhood-based Similarity)&lt;/code&gt;는 &lt;strong&gt;두 정점이 많은 이웃을 공유할수록 유사&lt;/strong&gt;하다고 간주한다. 공유하는 이웃의 수에 비례하여 유사도가 높아진다.&lt;/p&gt;&lt;p&gt;정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;,&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 이웃집합을 각각 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N(u)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N(v)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이라고 하자. 두 정점의 공통 이웃 수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;S_{u,v}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.969438em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.05764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 다음과 같이 정의한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∩&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∩&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;S_{u,v} = |N(u)\cap N(v)| = \sum_{w\in N(u)\cap N(v)} 1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.969438em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.05764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∩&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.5660100000000003em;vertical-align:-1.516005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;∩&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이 때 손실함수는 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msub&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathcal{L} = \sum_{(u,v)\in V\times V}\Vert z_u^{\top}z_v-\textcolor{red}{S_{u,v}}\Vert^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.5660100000000003em;vertical-align:-1.516005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.150216em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em;color:red&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.05764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot; style=&quot;color:red&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:red&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;단순히 공통 이웃 수를 세는 것이 아니라, &lt;strong&gt;자카드 유사도&lt;/strong&gt;나 &lt;strong&gt;Adamic Adar 점수&lt;/strong&gt;를 사용할 수도 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;&lt;code&gt;자카드 유사도(Jaccard Similarity)&lt;/code&gt; : 공통 이웃의 수 대신 비율을 계산하는 방식&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo&gt;≤&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∩&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∪&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;≤&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;0\le \frac{N(u)\cap N(v)}{N(u)\cup N(v)} \le 1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.78041em;vertical-align:-0.13597em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≤&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.363em;vertical-align:-0.936em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∪&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∩&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≤&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;code&gt;Adamic Adar 점수&lt;/code&gt; : 공통 이웃 각각에 가중치를 부여하여 가중합을 계산하는 방식. 연결성(Degree)이 높은 노드일수록 가중치가 적다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∩&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;/msub&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sum_{w\in N(u)\cap N(v)}\frac{1}{d_w}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.8374449999999998em;vertical-align:-1.516005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;∩&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.3139999999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8360000000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;임의보행-기반-유사도&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9E%84%EC%9D%98%EB%B3%B4%ED%96%89-%EA%B8%B0%EB%B0%98-%EC%9C%A0%EC%82%AC%EB%8F%84&quot; aria-label=&quot;임의보행 기반 유사도 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;임의보행 기반 유사도&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;임의보행 기반 유사도(Random walk-based similarity)&lt;/code&gt;&lt;/strong&gt;는 &lt;strong&gt;한 정점에서 시작하여 랜덤 워크를 할 때 다른 정점에 도달할 확률&lt;/strong&gt;을 유사도로 간주한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:998px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:91.40624999999999%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;random-walk&quot; title=&quot;random-walk&quot; src=&quot;/static/c6fdcdb6a59ef528353650e85c58a293/350de/random-walk.png&quot; srcSet=&quot;/static/c6fdcdb6a59ef528353650e85c58a293/6f3f2/random-walk.png 256w,/static/c6fdcdb6a59ef528353650e85c58a293/01e7c/random-walk.png 512w,/static/c6fdcdb6a59ef528353650e85c58a293/350de/random-walk.png 998w&quot; sizes=&quot;(max-width: 998px) 100vw, 998px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;여기서 &lt;code&gt;랜덤워크&lt;/code&gt;란, 현재 정점의 이웃 중 하나를 균일한 확률로 선택하여 이동하는 과정을 반복하는 것을 의미한다. &lt;/p&gt;&lt;p&gt;이 경우 시작 정점 주변의 이웃 노드들, 각 노드의 연결성 등에 따라서 확률이 달라진다. 또, 거리기반 유사도처럼 유사도 측정의 최대거리 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 제한하지 않으므로, 그래프의 모든 노드들에 대하여 계산할 수 있게 된다.&lt;/p&gt;&lt;p&gt;따라서, &lt;strong&gt;지역적 정보&lt;/strong&gt;와 &lt;strong&gt;그래프 전역 정보&lt;/strong&gt;를 모두 고려한다는 장점이 있다.&lt;/p&gt;&lt;h3 id=&quot;과정&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B3%BC%EC%A0%95&quot; aria-label=&quot;과정 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;과정&lt;/h3&gt;&lt;ol&gt;&lt;li&gt;&lt;p&gt;각 정점에서 시작하여 랜덤 워크를 반복 수행한다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;각 정점에서 시작한 랜덤 워크 중 도달한 정점들의 리스트를 구성한다. &lt;strong&gt;정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서 시작하여 랜덤 워크 중 도달한 정점들의 리스트를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N_R(u)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10903em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.00773em&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/strong&gt;라고 한다. 한 정점을 여러번 도달한 경우, 해당 정점은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;**N_R(u)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.46528em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∗&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∗&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10903em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.00773em&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 여러번 포함될 수 있다**.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;다음의 손실함수를 최소화하는 임베딩을 학습한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mstyle&gt;&lt;mstyle mathcolor=&quot;orange&quot;&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mstyle&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathcal{L} = \textcolor{green}{\sum_{u\in V}}\textcolor{orange}{\sum_{v\in N_R(u)}} - \log(\textcolor{blue}{P(v|z_u)})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.5660100000000003em;vertical-align:-1.516005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.8556639999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:green&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em;color:green&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;color:green&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.321706em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:orange&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:orange&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em;color:orange&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3567071428571427em;margin-left:-0.10903em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.00773em;color:orange&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.14329285714285717em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot; style=&quot;color:orange&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:orange&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot; style=&quot;color:orange&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;color:orange&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:blue&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:blue&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:blue&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em;color:blue&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:blue&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{blue}{P(v|z_u)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:blue&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:blue&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:blue&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em;color:blue&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:blue&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서 시작한 랜덤워크가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 도달할 확률을 임베딩으로부터 추정한 결과를 의미한다. 이 (확률)값이 크면 클수록 좋다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;exp&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi&gt;exp&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{blue}{P(v|z_u)} = \frac{\exp(z_u^\top z_v)}{\sum_{n\in V}\exp(z_u^\top z_n)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:blue&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:blue&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:blue&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em;color:blue&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:blue&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.5391880000000002em;vertical-align:-1.01308em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.526108em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.17862099999999992em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32708000000000004em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;exp&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7751079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop&quot;&gt;exp&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.849108em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.01308em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;즉, 유사도 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;exp&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\exp(z_u^\top z_v)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.099108em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;exp&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.849108em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 높을 수록 도달 확률이 높다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;orange&quot;&gt;&lt;msub&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{orange}{\sum_{v\in N_R(u)}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.22471em;vertical-align:-0.47471em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;color:orange;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.22528999999999993em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:orange&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:orange&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em;color:orange&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3567071428571427em;margin-left:-0.10903em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.00773em;color:orange&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.14329285714285717em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot; style=&quot;color:orange&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:orange&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot; style=&quot;color:orange&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.47471em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 랜덤 워크 중 실제로 마주친(도달한) 모든 정점에 대하여 합산&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;msub&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{green}{\sum_{u\in V}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.07708em;vertical-align:-0.32708000000000004em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;color:green;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.17862099999999992em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:green&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em;color:green&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32708000000000004em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 모든 시작점마다 수행하여 합산&lt;/p&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;deepwalk와-node2vec&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#deepwalk%EC%99%80-node2vec&quot; aria-label=&quot;deepwalk와 node2vec permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;DeepWalk와 Node2Vec&lt;/h3&gt;&lt;p&gt;랜덤워크의 방식에 따라 &lt;strong&gt;&lt;code&gt;DeepWalk&lt;/code&gt;&lt;/strong&gt;와 &lt;strong&gt;&lt;code&gt;Node2Vec&lt;/code&gt;&lt;/strong&gt;으로 나눌 수 있다.&lt;/p&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;&lt;code&gt;DeepWalk&lt;/code&gt;는 위에서 설명했던 것과 같이 균일한 확률의 랜덤워크를 수행한다.&lt;/p&gt;&lt;p&gt;&lt;code&gt;Node2Vec&lt;/code&gt;은  &lt;strong&gt;&lt;code&gt;2차 치우친 임의보행(Second-order Biased Random Walk)&lt;/code&gt;&lt;/strong&gt;를 사용한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:52.34375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;2nd-order-biased&quot; title=&quot;2nd-order-biased&quot; src=&quot;/static/e874d974a8ae0511038cc9e5ddfa932c/2bef9/2nd-order-biased.png&quot; srcSet=&quot;/static/e874d974a8ae0511038cc9e5ddfa932c/6f3f2/2nd-order-biased.png 256w,/static/e874d974a8ae0511038cc9e5ddfa932c/01e7c/2nd-order-biased.png 512w,/static/e874d974a8ae0511038cc9e5ddfa932c/2bef9/2nd-order-biased.png 1024w,/static/e874d974a8ae0511038cc9e5ddfa932c/bb543/2nd-order-biased.png 1418w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;strong&gt;현재 정점(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)&lt;/strong&gt;과 &lt;strong&gt;직전에 머물렀던 정점(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)&lt;/strong&gt;을 모두 고려하여 다음 정점을 선택하며, &lt;strong&gt;직전 정점의 거리를 기준으로 경우(이전의 정점과 거리가 가까워지는/유지되는/멀어지는 경우)를 구분하여 차등적인 확률을 부여&lt;/strong&gt;한다. &lt;/p&gt;&lt;p&gt;이 차등적인 확률은 하이퍼파라미터이며, 어떻게 부여했는지에 따라 다른 임베딩을 얻는다. 아래의 그림은 Node2Vec 방식으로 K-means 군집 분석을 수행한 결과이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:30.46875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;node2vec&quot; title=&quot;node2vec&quot; src=&quot;/static/da0b3fefd15fa35414bfcf4e0ddfe609/2bef9/node2vec.png&quot; srcSet=&quot;/static/da0b3fefd15fa35414bfcf4e0ddfe609/6f3f2/node2vec.png 256w,/static/da0b3fefd15fa35414bfcf4e0ddfe609/01e7c/node2vec.png 512w,/static/da0b3fefd15fa35414bfcf4e0ddfe609/2bef9/node2vec.png 1024w,/static/da0b3fefd15fa35414bfcf4e0ddfe609/71c1d/node2vec.png 1536w,/static/da0b3fefd15fa35414bfcf4e0ddfe609/a878e/node2vec.png 2048w,/static/da0b3fefd15fa35414bfcf4e0ddfe609/24488/node2vec.png 2984w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;멀어지는 방향에 높은 확률을 부여하면, 정점의 &amp;#x27;역할&amp;#x27;(bridge 역할, 변두리 정점 등)이 같은 경우 임베딩이 유사해진다.&lt;/li&gt;&lt;li&gt;가까워지는 방향에 높은 확률을 부여하면, 같은 군집에 속한 정점들끼리 임베딩이 유사하게 나온다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;손실-함수-근사&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%86%90%EC%8B%A4-%ED%95%A8%EC%88%98-%EA%B7%BC%EC%82%AC&quot; aria-label=&quot;손실 함수 근사 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;손실 함수 근사&lt;/h3&gt;&lt;p&gt;랜덤 워크 기반의 손실함수는 정점의 수의 제곱에 비례하는 시간이 소요된다. 모든 시작점 각각에 대하여, 모든 도달점의 경우를 고려하기 때문이다.&lt;/p&gt;&lt;p&gt;이런 제곱합을 피하기 위해서 대부분 근사식을 사용하는데, 모든 정점에 대해 정규화하는 대신 &lt;strong&gt;몇개의 정점을 뽑아 비교하는 형태&lt;/strong&gt;이다. 이 때 뽑힌 정점들을 Negative sample이라고 부른다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;exp&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi&gt;exp&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;≈&lt;/mo&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/munderover&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mtext&gt;확률분포&lt;/mtext&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
\log\bigg(\frac{\exp(z_u^\top z_v)}{\sum_{n\in V}\exp(z_u^\top z_n)}\bigg)&amp;amp;\approx log(\sigma(z_u^\top z_n)) - \sum_{i=1}^k\log(\sigma(z_u^\top z_n)),n_i\\ &amp;amp;\sim P_V(확률분포)
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:4.913782em;vertical-align:-2.206891em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.706891em&quot;&gt;&lt;span style=&quot;top:-4.706891em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.836113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.526108em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.17862099999999992em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32708000000000004em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;exp&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7751079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop&quot;&gt;exp&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.849108em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.01308em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.289222em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.836113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.206891em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.706891em&quot;&gt;&lt;span style=&quot;top:-4.706891em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.836113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.8361130000000003em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.289222em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.836113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;확&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;률&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;분&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;포&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.206891em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sigma&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : sigmoid 함수&lt;/li&gt;&lt;li&gt;Negative sample의 숫자가 더 많을수록 학습이 안정적이다.&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;변환식-노드-임베딩과-귀납식-노드-임베딩&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%B3%80%ED%99%98%EC%8B%9D-%EB%85%B8%EB%93%9C-%EC%9E%84%EB%B2%A0%EB%94%A9%EA%B3%BC-%EA%B7%80%EB%82%A9%EC%8B%9D-%EB%85%B8%EB%93%9C-%EC%9E%84%EB%B2%A0%EB%94%A9&quot; aria-label=&quot;변환식 노드 임베딩과 귀납식 노드 임베딩 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;변환식 노드 임베딩과 귀납식 노드 임베딩&lt;/h2&gt;&lt;p&gt;지금까지 언급한 노드 임베딩 방식들은 &lt;strong&gt;&lt;code&gt;변환식(Transductive) 방법&lt;/code&gt;&lt;/strong&gt;으로, 학습의 결과로 정점의 임베딩 값 자체를 출력값으로 얻는다.&lt;/p&gt;&lt;p&gt;그런데, 이런 변환식 방법에는 여러 한계가 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;학습이 진행된 이후에 추가된 정점에 대해서는 임베딩을 얻을 수 없다. 다시 학습을 수행해야한다.&lt;/li&gt;&lt;li&gt;모든 정점에 대한 임베딩을 미리 계산하여 저장해두고 꺼내 써야한다.&lt;/li&gt;&lt;li&gt;정점이 속성(Attribute) 정보를 가진 경우 이를 활용할 수 없다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;따라서 이런 한계점을 극복한 노드 임베딩 방식으로, 정점을 임베딩으로 변화시키는 &amp;#x27;함수&amp;#x27;, 즉 &lt;strong&gt;인코더&lt;/strong&gt;를 얻는 &lt;strong&gt;&lt;code&gt;귀납식(Inductive) 방법&lt;/code&gt;&lt;/strong&gt;이 나오게 되었다. 귀납식 노드 임베딩 방식의 대표적인 방식이 &lt;strong&gt;&lt;code&gt;그래프 신경망(Graph Neural Network)&lt;/code&gt;&lt;/strong&gt;이다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[Graph 06 - 추천 시스템]]></title><description><![CDATA[그래프를 추천시스템에 어떻게 활용할까 by 신기정 교수님, BoostCamp AI Tech 5주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/35_recommender_system/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/35_recommender_system/</guid><pubDate>Thu, 25 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;추천시스템-기초&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%B6%94%EC%B2%9C%EC%8B%9C%EC%8A%A4%ED%85%9C-%EA%B8%B0%EC%B4%88&quot; aria-label=&quot;추천시스템 기초 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;추천시스템 기초&lt;/h1&gt;&lt;p&gt;아마존 고객 맞춤 상품 / 넷플릭스 영화 추천 / 유튜브 영상 추천 / 페이스북 친구 추천 등&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;추천 시스템&lt;/code&gt;&lt;/strong&gt;은 사용자 각각이 구매/선호할만한 상품을 추천하는 기법이다.&lt;/p&gt;&lt;p&gt;사용자별 구매기록은 그래프로 표현 가능한데, 두가지로 나눌 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;암시적(Implicit) 선호 - ex)구매 기록&lt;/li&gt;&lt;li&gt;명시적(Explicit) 선호 - ex)평점&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:800px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:97.65625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;recommend&quot; title=&quot;recommend&quot; src=&quot;/static/aad52cfdb8c6423c2f9f2ffa01c8aade/5a190/recommend.png&quot; srcSet=&quot;/static/aad52cfdb8c6423c2f9f2ffa01c8aade/6f3f2/recommend.png 256w,/static/aad52cfdb8c6423c2f9f2ffa01c8aade/01e7c/recommend.png 512w,/static/aad52cfdb8c6423c2f9f2ffa01c8aade/5a190/recommend.png 800w&quot; sizes=&quot;(max-width: 800px) 100vw, 800px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;추천 시스템의 핵심은 사용자별 구매를 예측하거나, 선호를 추정하는 것인데, 그래프 관점에서 이를 풀었을 때 &lt;strong&gt;&amp;quot;미래의 간선을 예측하는 문제&amp;quot;&lt;/strong&gt;혹은 &lt;strong&gt;&amp;quot;누락된 간선의 가중치(선호)를 추정하는 문제&amp;quot;&lt;/strong&gt;로 해석할 수 있다.&lt;/p&gt;&lt;h2 id=&quot;내용-기반-추천시스템&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%82%B4%EC%9A%A9-%EA%B8%B0%EB%B0%98-%EC%B6%94%EC%B2%9C%EC%8B%9C%EC%8A%A4%ED%85%9C&quot; aria-label=&quot;내용 기반 추천시스템 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;내용 기반 추천시스템&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;내용 기반 추천시스템&lt;/code&gt;&lt;/strong&gt;은 각 사용자가 구매/만족했던 상품과 유사한 것을 추천하는 방식이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;동일한 장르의 영화&lt;/li&gt;&lt;li&gt;동일한 감독의 영화 혹은 동일 배우가 출연한 영화&lt;/li&gt;&lt;li&gt;동일한 카테고리의 상품&lt;/li&gt;&lt;li&gt;동갑의 같은 학교를 졸업한 사람&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;원리&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9B%90%EB%A6%AC&quot; aria-label=&quot;원리 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;원리&lt;/h3&gt;&lt;ol&gt;&lt;li&gt;사용자가 선호했던 상품들의 &lt;code&gt;상품 프로필(Item Profile)&lt;/code&gt;을 수집한다.&lt;ul&gt;&lt;li&gt;상품 프로필이란, 해당 상품의 특성을 나열한 벡터이다.&lt;/li&gt;&lt;li&gt;영화의 경우 감독/장르/배우 등의 &lt;strong&gt;원-핫 인코딩&lt;/strong&gt;이 상품 프로필로 사용될 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;사용자가 선호한 상품 프로필의 선호도를 가중평균하여 &lt;code&gt;사용자 프로필&lt;/code&gt;을 구성한다.&lt;ul&gt;&lt;li&gt;즉, 사용자 프로필 역시 벡터이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;사용자 프로필 벡터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;true&quot;&gt;⇀&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\overrightharpoon{u}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.9525600000000001em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9525600000000001em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3.43056em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail&quot; style=&quot;height:0.522em;min-width:0.888em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.522em&quot; viewBox=&quot;0 0 400000 522&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M0 241v40h399993c4.7-4.7 7-9.3 7-14 0-9.3
-3.7-15.3-11-18-92.7-56.7-159-133.7-199-231-3.3-9.3-6-14.7-8-16-2-1.3-7-2-15-2
-10.7 0-16.7 2-18 6-2 2.7-1 9.7 3 21 15.3 42 36.7 81.8 64 119.5 27.3 37.7 58
 69.2 92 94.5zm0 0v40h399900v-40z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 (추천할만한 다른) 상품 프로필 벡터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;true&quot;&gt;⇀&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\overrightharpoon{v}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.9525600000000001em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9525600000000001em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3.43056em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail&quot; style=&quot;height:0.522em;min-width:0.888em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.522em&quot; viewBox=&quot;0 0 400000 522&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M0 241v40h399993c4.7-4.7 7-9.3 7-14 0-9.3
-3.7-15.3-11-18-92.7-56.7-159-133.7-199-231-3.3-9.3-6-14.7-8-16-2-1.3-7-2-15-2
-10.7 0-16.7 2-18 6-2 2.7-1 9.7 3 21 15.3 42 36.7 81.8 64 119.5 27.3 37.7 58
 69.2 92 94.5zm0 0v40h399900v-40z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 코사인 유사도 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;true&quot;&gt;⇀&lt;/mo&gt;&lt;/mover&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;true&quot;&gt;⇀&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo stretchy=&quot;true&quot;&gt;⇀&lt;/mo&gt;&lt;/mover&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;true&quot;&gt;⇀&lt;/mo&gt;&lt;/mover&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{\overrightharpoon{u}\cdot\overrightharpoon{v}}{\Vert\overrightharpoon{u}\Vert\Vert\overrightharpoon{v}\Vert}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.712584em;vertical-align:-0.651792em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0607920000000002em&quot;&gt;&lt;span style=&quot;top:-2.523208em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord accent mtight&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9525600000000001em&quot;&gt;&lt;span style=&quot;top:-2.7em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3.13056em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail mtight&quot; style=&quot;height:0.522em;min-width:0.888em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.522em&quot; viewBox=&quot;0 0 400000 522&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M0 241v40h399993c4.7-4.7 7-9.3 7-14 0-9.3
-3.7-15.3-11-18-92.7-56.7-159-133.7-199-231-3.3-9.3-6-14.7-8-16-2-1.3-7-2-15-2
-10.7 0-16.7 2-18 6-2 2.7-1 9.7 3 21 15.3 42 36.7 81.8 64 119.5 27.3 37.7 58
 69.2 92 94.5zm0 0v40h399900v-40z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord accent mtight&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9525600000000001em&quot;&gt;&lt;span style=&quot;top:-2.7em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3.13056em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail mtight&quot; style=&quot;height:0.522em;min-width:0.888em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.522em&quot; viewBox=&quot;0 0 400000 522&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M0 241v40h399993c4.7-4.7 7-9.3 7-14 0-9.3
-3.7-15.3-11-18-92.7-56.7-159-133.7-199-231-3.3-9.3-6-14.7-8-16-2-1.3-7-2-15-2
-10.7 0-16.7 2-18 6-2 2.7-1 9.7 3 21 15.3 42 36.7 81.8 64 119.5 27.3 37.7 58
 69.2 92 94.5zm0 0v40h399900v-40z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∥&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.394em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord accent mtight&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9525600000000001em&quot;&gt;&lt;span style=&quot;top:-2.7em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3.13056em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail mtight&quot; style=&quot;height:0.522em;min-width:0.888em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.522em&quot; viewBox=&quot;0 0 400000 522&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M0 241v40h399993c4.7-4.7 7-9.3 7-14 0-9.3
-3.7-15.3-11-18-92.7-56.7-159-133.7-199-231-3.3-9.3-6-14.7-8-16-2-1.3-7-2-15-2
-10.7 0-16.7 2-18 6-2 2.7-1 9.7 3 21 15.3 42 36.7 81.8 64 119.5 27.3 37.7 58
 69.2 92 94.5zm0 0v40h399900v-40z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mord accent mtight&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9525600000000001em&quot;&gt;&lt;span style=&quot;top:-2.7em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3.13056em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail mtight&quot; style=&quot;height:0.522em;min-width:0.888em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.522em&quot; viewBox=&quot;0 0 400000 522&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M0 241v40h399993c4.7-4.7 7-9.3 7-14 0-9.3
-3.7-15.3-11-18-92.7-56.7-159-133.7-199-231-3.3-9.3-6-14.7-8-16-2-1.3-7-2-15-2
-10.7 0-16.7 2-18 6-2 2.7-1 9.7 3 21 15.3 42 36.7 81.8 64 119.5 27.3 37.7 58
 69.2 92 94.5zm0 0v40h399900v-40z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.651792em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 계산하여 매칭한다.&lt;ul&gt;&lt;li&gt;즉, 두 벡터의 사이각의 코사인 값을 계산한다.(두 벡터의 내적값 / 두 벡터의 크기의 곱)&lt;ul&gt;&lt;li&gt;두 벡터가 비슷한 방향을 바라보고 있거나, 크기가 비슷할수록 코사인 유사도가 높게 나온다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;코사인 유사도가 높을수록, 해당 사용자가 선호/구매했던 물품들과 비슷한 상품이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;계산된 코사인 유사도가 가장 높은 상품을 사용자에게 추천한다.&lt;/li&gt;&lt;/ol&gt;&lt;h3 id=&quot;장단점&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9E%A5%EB%8B%A8%EC%A0%90&quot; aria-label=&quot;장단점 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;장/단점&lt;/h3&gt;&lt;table&gt;&lt;thead&gt;&lt;tr&gt;&lt;th&gt;장점&lt;/th&gt;&lt;th&gt;단점&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;&lt;tbody&gt;&lt;tr&gt;&lt;td&gt;다른 사용자의 구매기록이 필요없음&lt;/td&gt;&lt;td&gt;부가정보가 없는 경우 사용 불가&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;독특한 취향의 사용자에게도 추천 가능&lt;/td&gt;&lt;td&gt;구매기록이 없는 사용자에게는 사용 불가&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;새 상품에 대한 추천 가능&lt;/td&gt;&lt;td&gt;Overfitting으로 지나치게 협소한 추천을 할 위험이 있음&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;추천의 이유를 제공할 수 있음&lt;/td&gt;&lt;td&gt;&lt;/td&gt;&lt;/tr&gt;&lt;/tbody&gt;&lt;/table&gt;&lt;h2 id=&quot;협업-필터링-추천-시스템&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%98%91%EC%97%85-%ED%95%84%ED%84%B0%EB%A7%81-%EC%B6%94%EC%B2%9C-%EC%8B%9C%EC%8A%A4%ED%85%9C&quot; aria-label=&quot;협업 필터링 추천 시스템 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;협업 필터링 추천 시스템&lt;/h2&gt;&lt;p&gt;내용 기반 추천시스템을 일부 보완하는 추천방식이다.&lt;/p&gt;&lt;h3 id=&quot;원리-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9B%90%EB%A6%AC-1&quot; aria-label=&quot;원리 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;원리&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;사용자-사용자 협업 필터링&lt;/code&gt;&lt;/strong&gt;이란, 예상 사용자에게 유사한 취향을 가진 다른 사용자들이 사용했던 상품을 추천하는 것이다.&lt;/p&gt;&lt;p&gt; 추천의 대상 사용자를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 할 때,&lt;/p&gt;&lt;ol&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 유사한 취향의 사용자들을 찾는다.&lt;/li&gt;&lt;li&gt;유사한 취향의 사용자들이 선호한 상품을 찾는다.&lt;/li&gt;&lt;li&gt;이 상품들을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에게 추천한다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;이 때 &lt;strong&gt;취향의 유사도&lt;/strong&gt;는 &lt;strong&gt;&lt;code&gt;상관계수(Correlation Coefficient)&lt;/code&gt;&lt;/strong&gt;를 통해 측정한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;true&quot;&gt;‾&lt;/mo&gt;&lt;/mover&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;true&quot;&gt;‾&lt;/mo&gt;&lt;/mover&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msqrt&gt;&lt;mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;true&quot;&gt;‾&lt;/mo&gt;&lt;/mover&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/msqrt&gt;&lt;msqrt&gt;&lt;mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;true&quot;&gt;‾&lt;/mo&gt;&lt;/mover&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/msqrt&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;sim(x,y) = \frac{\sum_{s\in S_{xy}}(r_{xs}-\overline{r_x})(r_{ys}-\overline{r_y})}{\sqrt{\sum_{s\in S_{xy}}(r_{xs}-\overline{r_x})^2}\sqrt{\sum_{s\in S_{xy}}(r_{ys}-\overline{r_y})^2}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.36703em;vertical-align:-1.7300000000000002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.63703em&quot;&gt;&lt;span style=&quot;top:-2.11em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.111485em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord sqrt&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.111485em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3.8em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.8em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;padding-left:1em&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.17862099999999997em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.16454285714285716em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.05764em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2818857142857143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.49703em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord overline&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.63056em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.55056em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;overline-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.740108em&quot;&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.071485em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.8em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail&quot; style=&quot;min-width:1.02em;height:1.8800000000000001em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;1.8800000000000001em&quot; viewBox=&quot;0 0 400000 1944&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M983 90
l0 -0
c4,-6.7,10,-10,18,-10 H400000v40
H1013.1s-83.4,268,-264.1,840c-180.7,572,-277,876.3,-289,913c-4.7,4.7,-12.7,7,-24,7
s-12,0,-12,0c-1.3,-3.3,-3.7,-11.7,-7,-25c-35.3,-125.3,-106.7,-373.3,-214,-744
c-10,12,-21,25,-33,39s-32,39,-32,39c-6,-5.3,-15,-14,-27,-26s25,-30,25,-30
c26.7,-32.7,52,-63,76,-91s52,-60,52,-60s208,722,208,722
c56,-175.3,126.3,-397.3,211,-666c84.7,-268.7,153.8,-488.2,207.5,-658.5
c53.7,-170.3,84.5,-266.8,92.5,-289.5z
M1001 80h400000v40h-400000z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.728515em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord sqrt&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.111485em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3.8em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.8em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;padding-left:1em&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.17862099999999997em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.16454285714285716em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.05764em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2818857142857143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.49703em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord overline&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.63056em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.55056em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;overline-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.740108em&quot;&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.071485em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.8em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail&quot; style=&quot;min-width:1.02em;height:1.8800000000000001em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;1.8800000000000001em&quot; viewBox=&quot;0 0 400000 1944&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M983 90
l0 -0
c4,-6.7,10,-10,18,-10 H400000v40
H1013.1s-83.4,268,-264.1,840c-180.7,572,-277,876.3,-289,913c-4.7,4.7,-12.7,7,-24,7
s-12,0,-12,0c-1.3,-3.3,-3.7,-11.7,-7,-25c-35.3,-125.3,-106.7,-373.3,-214,-744
c-10,12,-21,25,-33,39s-32,39,-32,39c-6,-5.3,-15,-14,-27,-26s25,-30,25,-30
c26.7,-32.7,52,-63,76,-91s52,-60,52,-60s208,722,208,722
c56,-175.3,126.3,-397.3,211,-666c84.7,-268.7,153.8,-488.2,207.5,-658.5
c53.7,-170.3,84.5,-266.8,92.5,-289.5z
M1001 80h400000v40h-400000z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.728515em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.341485em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.111485em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.998515em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.111485em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.17862099999999997em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.16454285714285716em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.05764em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2818857142857143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.49703em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord overline&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.63056em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.55056em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;overline-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord overline&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.63056em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.55056em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;overline-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7300000000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r_{xs}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 사용자 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 상품 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;s&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해 매긴 평점&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;true&quot;&gt;‾&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\overline{r_x}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.78056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord overline&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.63056em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.55056em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;overline-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 사용자 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 매긴 평균 평점&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;S_{xy}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.969438em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.05764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 사용자 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 공동 구매한 상품들&lt;/li&gt;&lt;li&gt;분모는 유사도 값을 -1과 1사이로 정규화하는 목적으로 사용한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;구체적으로는 &lt;strong&gt;취향의 유사도를 가중치로 사용한 평점의 가중 평균&lt;/strong&gt;을 통해 평점을 추정한다.&lt;/p&gt;&lt;p&gt;사용자 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 상품 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;s&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대한 평점을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r_{xs}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 추정할 때, 상관 계수를 이용하여 상품 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;s&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 구매한 사용자 중 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 취향이 유사한 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;명의 사용자 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;;&lt;/mo&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N(x;s)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 뽑는다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;;&lt;/mo&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;;&lt;/mo&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\hat{r}_{xs} = \frac{\sum_{y\in N(x;s)}sim(x,y)\cdot r_{ys}}{\sum_{y\in N(x;s)}sim(x,y)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.19444em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.7754200000000004em;vertical-align:-1.1607100000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.61471em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.22528999999999993em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.47471em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.86471em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.22528999999999993em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.47471em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1607100000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;취향이 유사한 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;명이라도, 유사한 정도의 차이가 있기 때문에 가중치 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;sim(x,y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 곱해준다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r_{ys}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 유사한 취향의 사람 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 매긴 평점&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 방식을 통해, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 아직 구매하지 않은 상품들에 대해 평점을 각각 추정하고, 추정한 평점이 가장 높은 상품들을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에게 추천한다.&lt;/p&gt;&lt;h3 id=&quot;장단점-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9E%A5%EB%8B%A8%EC%A0%90-1&quot; aria-label=&quot;장단점 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;장/단점&lt;/h3&gt;&lt;table&gt;&lt;thead&gt;&lt;tr&gt;&lt;th&gt;장점&lt;/th&gt;&lt;th&gt;단점&lt;/th&gt;&lt;/tr&gt;&lt;/thead&gt;&lt;tbody&gt;&lt;tr&gt;&lt;td&gt;상품에 대한 부가 정보(속성 정보)가 없어도 사용 가능&lt;/td&gt;&lt;td&gt;충분한 수의 평점 데이터가 누적되어야함&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;&lt;/td&gt;&lt;td&gt;새 상품, 새 사용자에 대한 추천 불가&lt;/td&gt;&lt;/tr&gt;&lt;tr&gt;&lt;td&gt;&lt;/td&gt;&lt;td&gt;독특한 취향의 사용자에게 추천 불가(취향이 유사한 사람을 찾아내기 어려움)&lt;/td&gt;&lt;/tr&gt;&lt;/tbody&gt;&lt;/table&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;추천-시스템-평가&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%B6%94%EC%B2%9C-%EC%8B%9C%EC%8A%A4%ED%85%9C-%ED%8F%89%EA%B0%80&quot; aria-label=&quot;추천 시스템 평가 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;추천 시스템 평가&lt;/h2&gt;&lt;h3 id=&quot;데이터-분리&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%8D%B0%EC%9D%B4%ED%84%B0-%EB%B6%84%EB%A6%AC&quot; aria-label=&quot;데이터 분리 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;데이터 분리&lt;/h3&gt;&lt;p&gt;추천시스템의 정확도를 평가하기위해서, &lt;strong&gt;&lt;code&gt;데이터 분리&lt;/code&gt;&lt;/strong&gt; 기법을 사용한다.&lt;/p&gt;&lt;p&gt;데이터를 훈련(Training) 데이터와 평가(Test) 데이터로 분리하고, 훈련 데이터를 이용하여 가려진 평가 데이터의 평점을 추정한다. 이후, 평가 데이터의 추정 평점과 실제 평점을 비교하여 오차를 측정하고 학습한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때 오차 측정 지표로는 &lt;strong&gt;&lt;code&gt;MSE&lt;/code&gt;&lt;/strong&gt;가 많이 사용된다.&lt;/li&gt;&lt;li&gt;평가 데이터 내의 평점들의 집합을 T라고 할 때, MSE는 다음과 같이 구할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{1}{|T|}\sum_{r_{xi}\in T}(r_{xi}-\hat{r}_{xi})^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.7158759999999997em;vertical-align:-1.394436em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.855664em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.02778em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.394436em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1141079999999999em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.19444em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;MSE에 제곱근을 씌운 &lt;strong&gt;&lt;code&gt;RMSE&lt;/code&gt;&lt;/strong&gt;도 많이 사용된다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;평가지표&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%8F%89%EA%B0%80%EC%A7%80%ED%91%9C&quot; aria-label=&quot;평가지표 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;평가지표&lt;/h3&gt;&lt;p&gt;이외에도 다양한 평가지표가 활용된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;추정한 평점으로 순위를 매긴 후, 실제 평점으로 매긴 순위와 상관계수 계산&lt;/li&gt;&lt;li&gt;추천한 상품 중 실제 구매로 이루어진 것의 비율 측정&lt;/li&gt;&lt;li&gt;추천의 순서 혹은 다양성까지 고려하는 지표들도 사용&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;추천-시스템-심화&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%B6%94%EC%B2%9C-%EC%8B%9C%EC%8A%A4%ED%85%9C-%EC%8B%AC%ED%99%94&quot; aria-label=&quot;추천 시스템 심화 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;추천 시스템 심화&lt;/h1&gt;&lt;p&gt;이전까지는 노드 임베딩을 적용시키지 않았을 때의 시스템이었다.&lt;br/&gt;
노드임베딩 기법으로 기존의 그래프들을 벡터로 표시하여 좌표공간내로 옮길 수 있게 되자, ML의 도구들을 사용한 추천시스템의 비약적인 발전이 있었다.  &lt;/p&gt;&lt;p&gt;이러한 추천시스템 발전의 기념비적 이벤트라고 볼 수 있는 &lt;em&gt;넷플릭스 챌린지&lt;/em&gt;의 예를 들어 발전 과정을 확인해보자.&lt;/p&gt;&lt;h2 id=&quot;넷플릭스-챌린지&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%84%B7%ED%94%8C%EB%A6%AD%EC%8A%A4-%EC%B1%8C%EB%A6%B0%EC%A7%80&quot; aria-label=&quot;넷플릭스 챌린지 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;넷플릭스 챌린지&lt;/h2&gt;&lt;p&gt;2000년부터 2005년까지 수집한 사용자별 영화 평점 데이터를 사용하였다.&lt;/p&gt;&lt;p&gt;트레이닝 데이터는 48만명 사용자의 1만 8천개 영화에 대한 1억개의 평점으로, 테스트 데이터는 각 사용자의 최신 평점 280만개로 구성되어있다.&lt;/p&gt;&lt;p&gt;목표는 당시 넷플릭스의 추천 시스템의 성능을 10% 이상 향상시키는 것이었다(평균 제곱근 오차 0.9514→0.8563)&lt;/p&gt;&lt;p&gt;2006년부터 2009년까지 진행되었고, 2700개의 팀이 참여하였으며, 추천시스템의 성능이 비약적으로 발전하는 계기가 되었다.&lt;/p&gt;&lt;h2 id=&quot;잠재-인수-모형&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9E%A0%EC%9E%AC-%EC%9D%B8%EC%88%98-%EB%AA%A8%ED%98%95&quot; aria-label=&quot;잠재 인수 모형 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;잠재 인수 모형&lt;/h2&gt;&lt;p&gt;잠재 인수 모형(Latent Factor Model)은 넷플릭스 챌린지에서 제안되어 큰 성능 개선을 이루어낸 모델로, 지금까지도 자주 사용되는 모형이다. UV 분해(UV Decompostion)이라고도 한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;수학적 개념 SVD와 유사하여 SVD로도 부르곤 하지만, 정확히는 약간 차이가 있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Latent Factor Model의 핵심은 &lt;strong&gt;사용자와 상품을 벡터로 표현하는 것&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:68.359375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;lfm-embedding-example&quot; title=&quot;lfm-embedding-example&quot; src=&quot;/static/604bb1bb7f0116f7688a66ac84c5389b/2bef9/lfm-embedding-example.png&quot; srcSet=&quot;/static/604bb1bb7f0116f7688a66ac84c5389b/6f3f2/lfm-embedding-example.png 256w,/static/604bb1bb7f0116f7688a66ac84c5389b/01e7c/lfm-embedding-example.png 512w,/static/604bb1bb7f0116f7688a66ac84c5389b/2bef9/lfm-embedding-example.png 1024w,/static/604bb1bb7f0116f7688a66ac84c5389b/71c1d/lfm-embedding-example.png 1536w,/static/604bb1bb7f0116f7688a66ac84c5389b/a878e/lfm-embedding-example.png 2048w,/static/604bb1bb7f0116f7688a66ac84c5389b/820f7/lfm-embedding-example.png 2126w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위는 사용자와 영화를 임베딩한 예시이다. 2차원 공간에 영화를 장르와 성격에 따라 배치하고, 사용자의 취향도 분석하여 같은 평면 위에 배치했다. 사용자는 이 평면 내에서 본인의 위치와 가까운 위치에 있는 영화를 추천받게 된다.&lt;/p&gt;&lt;p&gt;그러나, 이 방식은 영화의 [진지하고 가벼운 정도]나, [얼마나 로맨스 장르에 가깝고 얼마나 액션 장르에 가까운지]등의 모호한 개념들을 기준으로 좌표가 설정되어야한다는 어려움이 있었다. 따라서, 잠재 인수 모형은 측정하기 어려운 고정된 인수 대신 효과적인 인수를 학습하는 것을 목표로 한다. 즉, 추천을 가장 정확하게 할 수 있는 인수를 찾아 학습하는 것이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때 학습한 인수, 즉 임베딩 공간의 축들을 &lt;strong&gt;&lt;code&gt;잠재 인수(Latent Factor)&lt;/code&gt;&lt;/strong&gt;라고 부른다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:69.921875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;lfm-embedding-example2&quot; title=&quot;lfm-embedding-example2&quot; src=&quot;/static/3ded21af59ca73682f8e44d68b530e81/2bef9/lfm-embedding-example2.png&quot; srcSet=&quot;/static/3ded21af59ca73682f8e44d68b530e81/6f3f2/lfm-embedding-example2.png 256w,/static/3ded21af59ca73682f8e44d68b530e81/01e7c/lfm-embedding-example2.png 512w,/static/3ded21af59ca73682f8e44d68b530e81/2bef9/lfm-embedding-example2.png 1024w,/static/3ded21af59ca73682f8e44d68b530e81/71c1d/lfm-embedding-example2.png 1536w,/static/3ded21af59ca73682f8e44d68b530e81/a878e/lfm-embedding-example2.png 2048w,/static/3ded21af59ca73682f8e44d68b530e81/410b4/lfm-embedding-example2.png 2102w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h3 id=&quot;손실함수&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%86%90%EC%8B%A4%ED%95%A8%EC%88%98&quot; aria-label=&quot;손실함수 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;손실함수&lt;/h3&gt;&lt;p&gt;그렇다면, 사용자와 상품을 어떻게 임베딩해야 추천시스템에 적용하기 좋도록 임베딩할 수 있는가?&lt;/p&gt;&lt;p&gt;&lt;strong&gt;사용자-상품 임베딩의 내적(Inner Product)이 평점과 최대한 유사&lt;/strong&gt;하도록 하는 것이다.&lt;/p&gt;&lt;p&gt;이를 행렬 차원에서 이미지로 살펴보자.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:24.609375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAFCAIAAADKYVtkAAAACXBIWXMAABYlAAAWJQFJUiTwAAABBElEQVQY022MMVKDQBiFuZh6AU9hlSKdhU2K2Kq141ksHQscTRgLQggBVvZnYdllgA0bSLILrqn9ijfz3rz5LKU6ymwOn3zxyojNuaP00AmROw7z1up00lqbVMOg6lrluTL1jNmtUqTu+jZ5v44W92E82ayueCuz7y8ymdCHR0pIEARAabpckvkcI4QxTpIEIQQAVlYx378h6BLQE+AZ4LsfSOIsSnNacG5OK9fNqyqzP9h0SowlTTNC4Ix1PPbbzUscXsTxM8DbNpgJUYtm53teFEdC7vZy3xqkbItCYtx2nfzrbd/3ltZ9XYVCFE1DmwYxVkI6GiNOMKvK8tCM4ziM//MLEtYJI2tkXt4AAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;lfm-matrix&quot; title=&quot;lfm-matrix&quot; src=&quot;/static/692899adb6ec52064efecf2b35e42dda/2bef9/lfm-matrix.png&quot; srcSet=&quot;/static/692899adb6ec52064efecf2b35e42dda/6f3f2/lfm-matrix.png 256w,/static/692899adb6ec52064efecf2b35e42dda/01e7c/lfm-matrix.png 512w,/static/692899adb6ec52064efecf2b35e42dda/2bef9/lfm-matrix.png 1024w,/static/692899adb6ec52064efecf2b35e42dda/71c1d/lfm-matrix.png 1536w,/static/692899adb6ec52064efecf2b35e42dda/a878e/lfm-matrix.png 2048w,/static/692899adb6ec52064efecf2b35e42dda/1ead5/lfm-matrix.png 3304w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;사용자들의 임베딩(벡터)을 쌓아서 만든 사용자 행렬을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 한다. 위의 이미지에서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P^\top&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.849108em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.849108em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 각 열(세로벡터)은 사용자 한명의 임베딩을 의미한다.&lt;/li&gt;&lt;li&gt;영화들의 임베딩(벡터)을 쌓아서 만든 상품 행렬을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Q&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 한다. 위의 이미지에서 Q의 각 행(가로벡터)은 상품 하나의 임베딩을 의미한다.&lt;/li&gt;&lt;li&gt;사용자 수의 열과 상품 수의 행을 가진 평점 행렬을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;R&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.00773em&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;
이라고 하고, 이는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Q \cdot P^\top&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.849108em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.849108em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 같다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;사용자 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 임베딩을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, 상품 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 임베딩을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;q_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 하고, 사용자 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 상품 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대한 평점을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r_{xi}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 하자. 임베딩의 목표는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_x^\top q_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.096108em;vertical-align:-0.247em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.849108em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r_{xi}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 유사하도록 하는 것이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathcal{L} = \sum_{(i,x)\in R}(r_{xi}-p_x^\top q_i)^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.5660100000000003em;vertical-align:-1.516005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.00773em&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1491079999999998em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;훈련 데이터(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;R&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.00773em&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)에 있는 평점에 대해서만 계산한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;그러나 위 손실함수를 사용할 경우에는 일반적인 기계학습과 마찬가지로 Overfitting이 발생할 수 있다.  훈련데이터에만 fitting된 모델이 나올 수 있는 것이다.&lt;/p&gt;&lt;p&gt;이를 피하기 위하여 정규화 항을 손실 함수에 더해준다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;munder&gt;&lt;mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mtext&gt;오차&lt;/mtext&gt;&lt;/munder&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;munder&gt;&lt;munder&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;msub&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/munder&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;msub&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/munder&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mrow&gt;&lt;mtext&gt;모형복잡도&lt;/mtext&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mtext&gt;정규화항&lt;/mtext&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathcal{L} = \underbrace{ \sum_{(i,x)\in R}(r_{xi}-p_x^\top q_i)^2}_{오차} + \textcolor{red}{\underbrace{\textcolor{green}{\lambda_1}\sum_x\Vert p_x \Vert^2 + \textcolor{green}{\lambda_2}\sum_i\Vert q_i \Vert ^2}_{모형복잡도(정규화항)}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.8923410000000005em;vertical-align:-2.8423360000000004em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-0.20766899999999966em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.050005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;오&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;차&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.050005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-0.8859999999999997em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.050005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.050005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.00773em&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.1640050000000004em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.8423360000000004em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.875674em;vertical-align:-2.825669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500049999999999em&quot;&gt;&lt;span style=&quot;top:-0.3993360000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.050005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:red&quot;&gt;모&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:red&quot;&gt;형&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:red&quot;&gt;복&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:red&quot;&gt;잡&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:red&quot;&gt;도&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:red&quot;&gt;정&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:red&quot;&gt;규&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:red&quot;&gt;화&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:red&quot;&gt;항&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.050005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-1.1243360000000004em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;color:red;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;color:red;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;color:red;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;color:red;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500050000000005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;λ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;color:red;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.8999949999999999em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;color:red&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.250005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;color:red;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;color:red;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot; style=&quot;color:red&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;color:red;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;λ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;color:red;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000003em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;color:red&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;color:red;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:red&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.925669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.825669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;정규화는 손실함수를 학습할 때, 훈련데이터와의 오차만 줄이는 것이 아니라, 임베딩 값 자체도 줄이게 만든다. 임베딩 값이 너무 크면, 모델이 훈련데이터의 noise까지 완전히 학습하기 때문이다. 이렇게 모형복잡도도 최소화하여 일반화 성능을 높인다.&lt;ul&gt;&lt;li&gt;이는 위의 임베딩 좌표 상에서 볼 때, 각 좌표들의 크기를 줄여서 벡터들이 좀 더 모여있게 하는 결과를 가져온다. 즉, 기존에는 나와 아주 가까운 영화가 아니면 추천받지 못했지만, 정규화를 하고 나면 거리가 고만고만해지기 때문에 나와 비교적 조금 가까운 영화도 추천받을 수 있는 것이다.&lt;/li&gt;&lt;li&gt;잘 이해가 되지 않는다면, 배틀그라운드에서 에란겔과 사녹을 생각하자. 에란겔은 맵이 넓기 때문에 나와 아주 가까운 상대가 아니면 마주칠 일이 적다. 그러나, 사녹은 모두가 가까이 모여있기 때문에 조금만 가까이 있어도 마주치게 될 가능성이 높아지게 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;msub&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{green}{\lambda_1, \lambda_2}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;λ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:green&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;λ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 정규화의 세기를 의미하며, 하이퍼파라미터이다. 오차와 모형복잡도 중 어느것을 더 중점적으로 최소화할지 정해주는 역할을 한다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;최적화&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%B5%9C%EC%A0%81%ED%99%94&quot; aria-label=&quot;최적화 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;최적화&lt;/h3&gt;&lt;p&gt;손실함수를 최소화(Optimize)하는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Q&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 찾기 위해서 (확률적) 경사하강법을 사용한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;경사하강법은 손실함수를 안정적이지만 느리게 감소시킨다.&lt;/li&gt;&lt;li&gt;확률적 경사하강법은 손실함수를 불안정하지만 빠르게 감소시킨다.&lt;/li&gt;&lt;li&gt;실제로는 확률적 경사하강법이 더 많이 사용된다.&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;고급-잠재-인수-모형&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B3%A0%EA%B8%89-%EC%9E%A0%EC%9E%AC-%EC%9D%B8%EC%88%98-%EB%AA%A8%ED%98%95&quot; aria-label=&quot;고급 잠재 인수 모형 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;고급 잠재 인수 모형&lt;/h2&gt;&lt;p&gt;잠재 인수모형은 추천시스템의 성능을 크게 개선했지만, 넷플릭스 챌린지의 목표인 RMSE 0.8563에는 도달하지 못했다. 따라서 이를 추가적으로 더 개선시키기 위해 몇가지 테크닉을 더한 잠재 인수 모형이 나오게 되었다.&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;사용자와-상품의-편향을-고려한-잠재-인수-모형&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%82%AC%EC%9A%A9%EC%9E%90%EC%99%80-%EC%83%81%ED%92%88%EC%9D%98-%ED%8E%B8%ED%96%A5%EC%9D%84-%EA%B3%A0%EB%A0%A4%ED%95%9C-%EC%9E%A0%EC%9E%AC-%EC%9D%B8%EC%88%98-%EB%AA%A8%ED%98%95&quot; aria-label=&quot;사용자와 상품의 편향을 고려한 잠재 인수 모형 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;사용자와 상품의 편향을 고려한 잠재 인수 모형&lt;/h3&gt;&lt;p&gt;누군가 영화의 평점을 매길 때, 어떤 사람은 높은 점수를 잘 주는 경향이 있는가 하면, 어떤 사람은 눈이 까다로워 대부분의 점수를 낮게 주는 경우가 있다. 이런 사람간의 편차를 각 사용자의 &lt;strong&gt;&lt;code&gt;편향&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;/p&gt;&lt;p&gt;넷플릭스 챌린지에서, 각 사용자의 편향은 &lt;strong&gt;해당 사용자의 평점 평균과 전체 평점 평균의 차이&lt;/strong&gt;다. 해당 사용자의 평점 평균이 전체 평점 평균대비 높으면, 그 사용자는 후하게 점수를 주는 감이 있고, 오히려 낮으면, 그 사용자는 점수를 박하게 주는 편임을 알 수 있다. 이는 넷플릭스에서가 아닌 다른 상품의 추천시스템에서도 동일하게 적용된다.&lt;/p&gt;&lt;p&gt;따라서 개선된 잠재인수 모델에서는 평점을 &lt;strong&gt;&lt;code&gt;전체 평균&lt;/code&gt;&lt;/strong&gt;, &lt;strong&gt;&lt;code&gt;사용자 편향&lt;/code&gt;&lt;/strong&gt;, &lt;strong&gt;&lt;code&gt;상품 편향&lt;/code&gt;&lt;/strong&gt;, &lt;strong&gt;&lt;code&gt;상호작용&lt;/code&gt;&lt;/strong&gt;으로 분리한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mo&gt;&lt;mtext&gt;평점&lt;/mtext&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;/mo&gt;&lt;mtext&gt;전체평균&lt;/mtext&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;/mstyle&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/msub&gt;&lt;/mo&gt;&lt;mtext&gt;사용자편향&lt;/mtext&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;/mstyle&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mo&gt;&lt;mtext&gt;상품편향&lt;/mtext&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;/mstyle&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mstyle mathcolor=&quot;purple&quot;&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;msubsup&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mo&gt;&lt;mtext&gt;상호작용&lt;/mtext&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\underset{평점}{r_{xi}} = \textcolor{red}{\underset{전체평균}{\mu}} + \textcolor{blue}{\underset{사용자편향}{b_x}} + \textcolor{green}{\underset{상품편향}{b_i}} + \textcolor{purple}{\underset{상호작용}{p_x^\top q_i}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.324891em;vertical-align:-0.894331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.205669em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;평&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;점&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.894331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.5221010000000001em;vertical-align:-0.938771em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.161229em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:red&quot;&gt;전&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:red&quot;&gt;체&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:red&quot;&gt;평&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:red&quot;&gt;균&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop mathnormal&quot; style=&quot;color:red&quot;&gt;μ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938771em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.588771em;vertical-align:-0.894331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-2.205669em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:blue&quot;&gt;사&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:blue&quot;&gt;용&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:blue&quot;&gt;자&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:blue&quot;&gt;편&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:blue&quot;&gt;향&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.894331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.588771em;vertical-align:-0.894331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-2.205669em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:green&quot;&gt;상&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:green&quot;&gt;품&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:green&quot;&gt;편&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:green&quot;&gt;향&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.894331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.8904389999999998em;vertical-align:-0.9913310000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.108669em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:purple&quot;&gt;상&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:purple&quot;&gt;호&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:purple&quot;&gt;작&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot; style=&quot;color:purple&quot;&gt;용&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.9999999999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:purple&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:purple&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:purple&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9913310000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;기존에는 상호작용 항만을 가지고 평점을 예측하려고 했다면, 편향을 고려한 잠재인수 모형에서는 사용자와 상품의 편향, 전체 평균을 모두 고려한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 모형의 손실함수는 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/munder&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/munder&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/munder&gt;&lt;msubsup&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msubsup&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mn&gt;4&lt;/mn&gt;&lt;/msub&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/munder&gt;&lt;msubsup&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msubsup&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
\mathcal{L} = &amp;amp;\sum_{(i,x)\in R}(r_{xi}- (\mu + b_x + b_i + p_x^\top q_i))^2 \\
&amp;amp;+ [\lambda_1\sum_i\Vert p_x \Vert ^2 + \lambda_2\sum_i\Vert q_i \Vert ^2 + \lambda_3\sum_xb_x^2 + \lambda_4\sum_ib_i^2]
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:5.493684000000001em;vertical-align:-2.4968420000000004em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.9968420000000004em&quot;&gt;&lt;span style=&quot;top:-4.996842000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.130832em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.4968420000000004em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.9968420000000004em&quot;&gt;&lt;span style=&quot;top:-4.996842000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.00773em&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.130832em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;λ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000003em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;λ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000003em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;λ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.8999949999999999em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.250005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;λ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000003em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.4968420000000004em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이 모형을 적용시켜 RMSE가 0.89까지 개선되었다. 그러나 여전히 목표치에는 도달하기 어려웠다.&lt;/p&gt;&lt;h3 id=&quot;시간적-편향을-고려한-잠재인수-모형&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%8B%9C%EA%B0%84%EC%A0%81-%ED%8E%B8%ED%96%A5%EC%9D%84-%EA%B3%A0%EB%A0%A4%ED%95%9C-%EC%9E%A0%EC%9E%AC%EC%9D%B8%EC%88%98-%EB%AA%A8%ED%98%95&quot; aria-label=&quot;시간적 편향을 고려한 잠재인수 모형 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;시간적 편향을 고려한 잠재인수 모형&lt;/h3&gt;&lt;p&gt;넷플릭스의 사용자별 영화 평점 데이터가 수집되었던 2000년부터 2005년 구간 사이에,  시스템의 변화를 계기로 평점 평균이 크게 상승하는 일이 있었다. 또한, 영화의 평점은 출시일 이후 시간이 지남에 따라 상승하는 경향을 띤다. 상영이 끝난 작품은 입소문을 타서/좋아하는 감독의 작품이라서 등등 긍정적인 기대로 가지고 시청하는 재유입인원이 많기 때문이다.&lt;/p&gt;&lt;p&gt;이런 문제를 해소하기 위해 잠재인수 모형에 시간적 편향까지 고려하게 되었다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mo&gt;&lt;mtext&gt;평점&lt;/mtext&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;/mo&gt;&lt;mtext&gt;전체평균&lt;/mtext&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/msub&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mo&gt;&lt;mtext&gt;사용자편향&lt;/mtext&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mo&gt;&lt;mtext&gt;상품편향&lt;/mtext&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;msubsup&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;⊤&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mo&gt;&lt;mtext&gt;상호작용&lt;/mtext&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\underset{평점}{r_{xi}} = \underset{전체평균}{\mu} + \underset{사용자편향}{b_x\textcolor{red}{(t)}} + \underset{상품편향}{b_i\textcolor{red}{(t)}} + \underset{상호작용}{p_x^\top q_i}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.324891em;vertical-align:-0.894331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.205669em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;평&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;점&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.894331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.5221010000000001em;vertical-align:-0.938771em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.161229em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;전&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;체&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;평&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;균&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop mathnormal&quot;&gt;μ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938771em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.7443309999999999em;vertical-align:-0.994331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.75em&quot;&gt;&lt;span style=&quot;top:-2.105669em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;사&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;용&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;자&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;편&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;향&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.994331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.7443309999999999em;vertical-align:-0.994331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.75em&quot;&gt;&lt;span style=&quot;top:-2.105669em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;상&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;품&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;편&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;향&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.994331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.8904389999999998em;vertical-align:-0.9913310000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.108669em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;상&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;호&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;작&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;용&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.9999999999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8991079999999998em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;⊤&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9913310000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이 결과로 RMSE 0.876까지 도달하게 되었다.&lt;/p&gt;&lt;p&gt;이러한 획기적인 개선 이후에 넷플릭스 챌린지에서는 Ensemble 기법을 추가적으로 도입한 몇몇팀이 최종적으로 목표치에 도달하게 되었다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[Graph 04 - 군집 탐색]]></title><description><![CDATA[그래프의 구조를 어떻게 분석할까 by 신기정 교수님, BoostCamp AI Tech 5주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/33_community_detection/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/33_community_detection/</guid><pubDate>Wed, 24 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;군집탐색&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B5%B0%EC%A7%91%ED%83%90%EC%83%89&quot; aria-label=&quot;군집탐색 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;군집탐색&lt;/h1&gt;&lt;h3 id=&quot;실제-그래프에서의-군집-사례&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%8B%A4%EC%A0%9C-%EA%B7%B8%EB%9E%98%ED%94%84%EC%97%90%EC%84%9C%EC%9D%98-%EA%B5%B0%EC%A7%91-%EC%82%AC%EB%A1%80&quot; aria-label=&quot;실제 그래프에서의 군집 사례 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;실제 그래프에서의 군집 사례&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;온라인 소셜 네트워크&lt;ul&gt;&lt;li&gt;사회적 무리, 부정행위와 연관된 계정들, 분열된 조직체&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;키워드-광고주 그래프&lt;ul&gt;&lt;li&gt;동일 주제 키워드&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;뉴런간 연결 그래프&lt;ul&gt;&lt;li&gt;뇌의 기능적 구성 단위&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;그래프를 여러 군집으로 &amp;#x27;잘&amp;#x27;나누는 문제를 &lt;strong&gt;&lt;code&gt;군집 탐색(Community Detection)&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;/p&gt;&lt;h2 id=&quot;군집-구조의-통계적-유의성과-군집성&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B5%B0%EC%A7%91-%EA%B5%AC%EC%A1%B0%EC%9D%98-%ED%86%B5%EA%B3%84%EC%A0%81-%EC%9C%A0%EC%9D%98%EC%84%B1%EA%B3%BC-%EA%B5%B0%EC%A7%91%EC%84%B1&quot; aria-label=&quot;군집 구조의 통계적 유의성과 군집성 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;군집 구조의 통계적 유의성과 군집성&lt;/h2&gt;&lt;h3 id=&quot;배치모형&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%B0%B0%EC%B9%98%EB%AA%A8%ED%98%95&quot; aria-label=&quot;배치모형 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;배치모형&lt;/h3&gt;&lt;p&gt;군집 탐색이 성공적이었는지에 대한 판단 기준으로 &lt;strong&gt;&lt;code&gt;배치모형(Configuration Model)&lt;/code&gt;&lt;/strong&gt;을 사용한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:37.109375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAABJElEQVQY032RbUvCUBTH/SZ+Er9b9C6yIOxdL4LADKEItSgYS7NwqybLKHyYazq1pUnT3d3dPdx72nRhb/LPeeAczg/+cBLwK0ZZGPBHAVCf+fC/EmE6FlEe+7AE2aLljiGXH8CU42/Lp6I66RX2+A9mFDM3XaNXLoivfCeGv4fzy4OK9PCiVLUIv7uHZBK2tr+AtLTWQDEsB6myToC8y7qF0VD9nOpmDC/dikK9wgnR6BDA9srFGttscVE7k4Xnp/asWcoKIMmQSsHObrh3EZlZc9M3kYkDCGwT+4GPbcfF3gpu17SROrYRanAdamNIp2lmP9w3rpr5w5JkyEcb5xM2Ptm86I6062y1XnyLba9R4AbIwSgsyGXAiO1SSj3i+ST6wg+1xYCsYtR6iQAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;config-model&quot; title=&quot;config-model&quot; src=&quot;/static/9fb4394daf745e856a2df1aad9b6dba3/2bef9/config-model.png&quot; srcSet=&quot;/static/9fb4394daf745e856a2df1aad9b6dba3/6f3f2/config-model.png 256w,/static/9fb4394daf745e856a2df1aad9b6dba3/01e7c/config-model.png 512w,/static/9fb4394daf745e856a2df1aad9b6dba3/2bef9/config-model.png 1024w,/static/9fb4394daf745e856a2df1aad9b6dba3/c929c/config-model.png 1218w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;주어진 그래프에 대한 배치 모형은,&lt;/p&gt;&lt;ol&gt;&lt;li&gt;&lt;strong&gt;각 정점의 연결성(Degree)을 보존한 상태&lt;/strong&gt;에서&lt;/li&gt;&lt;li&gt;&lt;strong&gt;간선들을 무작위로 재배치&lt;/strong&gt;하여&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;얻은 그래프를 의미한다.&lt;/p&gt;&lt;p&gt;배치 모형에서 임의의 두 점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;j&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.85396em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 사이에 간선이 존재할 확률은 연결성에 비례한다.&lt;/p&gt;&lt;h3 id=&quot;군집성의-정의&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B5%B0%EC%A7%91%EC%84%B1%EC%9D%98-%EC%A0%95%EC%9D%98&quot; aria-label=&quot;군집성의 정의 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;군집성의 정의&lt;/h3&gt;&lt;p&gt;배치모형이 주어지면, 군집탐색의 성공 여부를 판단하기 위해 그래프와 배치모형의 &lt;strong&gt;&lt;code&gt;군집성(Modularity)&lt;/code&gt;&lt;/strong&gt;을 확인한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;#&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mtext&gt; in Graph&lt;/mtext&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;#&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mtext&gt; in ConfigModel&lt;/mtext&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{1}{2|E|}\sum_{s \in S}(\#edges\in s \text{\ in\ Graph}   - \mathbb{E}\#edges\in s \text{\ in\ ConfigModel})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.6431459999999998em;vertical-align:-1.321706em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.8556639999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.321706em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;#&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt; in Graph&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;#&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt; in ConfigModel&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;배치모형에서 기댓값을 사용하는 이유는, 배치모형의 간선의 수는 무작위이기 때문이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 때, 배치모형과 비교하여 &lt;strong&gt;그래프에서 군집 내부 간선의 수가 월등히 많을수록&lt;/strong&gt; 성공한 군집 탐색이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;무작위로 연결된 배치모형과의 비교를 통해 통계적 유의성을 판단한다.&lt;/li&gt;&lt;li&gt;정규화하기 때문에 항상 -1~1 사이의 값을 갖는다.&lt;/li&gt;&lt;li&gt;군집성이 일반적으로 0.3 ~ 0.7정도의 값을 가질때 그래프에 존재하는 통계적으로 유의미한 군집을 찾아냈다고 할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;군집-탐색-알고리즘&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B5%B0%EC%A7%91-%ED%83%90%EC%83%89-%EC%95%8C%EA%B3%A0%EB%A6%AC%EC%A6%98&quot; aria-label=&quot;군집 탐색 알고리즘 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;군집 탐색 알고리즘&lt;/h3&gt;&lt;h3 id=&quot;girvan-newman-알고리즘&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#girvan-newman-%EC%95%8C%EA%B3%A0%EB%A6%AC%EC%A6%98&quot; aria-label=&quot;girvan newman 알고리즘 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Girvan-Newman 알고리즘&lt;/h3&gt;&lt;p&gt;대표적인 하향식(Top-down) 군집 탐색 알고리즘으로, 전체그래프에서 군집들이 서로 분리되도록 &lt;strong&gt;서로 다른 군집간의 bridge 간선을 순차적으로 제거&lt;/strong&gt;한다.&lt;/p&gt;&lt;p&gt;이러한 다리 간선을 찾아내기 위하여, 간선의 &lt;strong&gt;&lt;code&gt;매개중심성(Betweenness Centrality)&lt;/code&gt;&lt;/strong&gt;을 이용한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;매개중심성이란, 해당 간선이 정점 간의 최단 경로에 놓이는 횟수를 의미한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로부터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;j&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.85396em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로의 최단 경로 수를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sigma_{i,j}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 하고, 그 중 간선 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(x,y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 포함한 것을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sigma_{i,j}(x,y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 할 때, 간선  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(x,y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 매개 중심성은 다음 수식으로 계산된다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;&amp;lt;&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;msub&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sum_{i&amp;lt;j}\frac{\sigma_{i,j}(x,y)}{\sigma_{i,j}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.840777em;vertical-align:-1.413777em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.8723309999999997em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;&amp;lt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.413777em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.972108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;그러므로, Girvan-Newman 알고리즘은 &lt;strong&gt;매개중심성이 높은 간선을 순차적으로 제거하여 군집을 분리&lt;/strong&gt;하는 방식이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;간선이 제거될 때마다, 매개 중심성을 다시 계산하여 갱신한다.&lt;/li&gt;&lt;li&gt;모든 간선이 제거될때까지 반복한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:35.9375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA2klEQVQY021QiW6DMBTj/7+xGxTKlftoWXPROTwmVdWQImE7tt9Ls22Ptm1zzq//PmsN1H3fP3himlKKENxZTRSx1mprKhNjXObJe3caDtVo6ZypZhwlWd9drFGUqgRb5tu6jAQFX4br1+PuTlVWla1TNSMPMSlFoyRnE1KhgfTO1DFKwVBQAbXiuGKUAElqbdZKcDajAStgMPz0/bf39q+Ko8dogSQYJF9vQ3c/1OZYLIxDG8KTbueU+usFJMHnzzYOHcoJxhCwBcHmNOT09qT7x+O/w2OVE/4CQzOVIcWhrqoAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;girvan-newman&quot; title=&quot;girvan-newman&quot; src=&quot;/static/c850c21348af9b8931acba5fec05042b/2bef9/girvan-newman.png&quot; srcSet=&quot;/static/c850c21348af9b8931acba5fec05042b/6f3f2/girvan-newman.png 256w,/static/c850c21348af9b8931acba5fec05042b/01e7c/girvan-newman.png 512w,/static/c850c21348af9b8931acba5fec05042b/2bef9/girvan-newman.png 1024w,/static/c850c21348af9b8931acba5fec05042b/71c1d/girvan-newman.png 1536w,/static/c850c21348af9b8931acba5fec05042b/064b2/girvan-newman.png 1874w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;이 때 간선의 제거 정도에 따라 다른 입도(Granularity)의 군집구조가 나타나는데, 군집성이 최대가 되는 지점까지 제거한다. 단, 현재 연결 요소들을 군집으로 가정하되 입력 그래프에서 군집성을 계산한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;군집성의 변화를 기록해두었다가, 최대가 되는 지점으로 복원한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:35.15625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAAAuklEQVQY041QXQ+CMBDz//8+HuQBEBMWRUDHDrZx+7BAxPCiNFnSbWnv2lOMMYSAMzHL+mGU8vEoThsz1oq8sH1vphlHxU3TZFmmiG6XkhXhxVprjHHO/RcTqbZtaRzu5dXrcf3w3psFPyy+a0/OVVketA5rCwtAsIXWGkHCBzsx7rB33reVMFKiOSJ6dt0LkBIc84dhUEqB7MSwROa6rnv0zBz8XDZSJElS5Hl6ToUQEDDz0iOceRO/AVnsmZK+GceJAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;girvan-newman2&quot; title=&quot;girvan-newman2&quot; src=&quot;/static/ac20653d0e3266934ed642f346e8a40c/2bef9/girvan-newman2.png&quot; srcSet=&quot;/static/ac20653d0e3266934ed642f346e8a40c/6f3f2/girvan-newman2.png 256w,/static/ac20653d0e3266934ed642f346e8a40c/01e7c/girvan-newman2.png 512w,/static/ac20653d0e3266934ed642f346e8a40c/2bef9/girvan-newman2.png 1024w,/static/ac20653d0e3266934ed642f346e8a40c/71c1d/girvan-newman2.png 1536w,/static/ac20653d0e3266934ed642f346e8a40c/09b15/girvan-newman2.png 1704w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h2 id=&quot;louvain-알고리즘&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#louvain-%EC%95%8C%EA%B3%A0%EB%A6%AC%EC%A6%98&quot; aria-label=&quot;louvain 알고리즘 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Louvain 알고리즘&lt;/h2&gt;&lt;p&gt;대표적인 상향식(Bottom-up) 군집 탐색 알고리즘으로, 개별 정점에서 시작하여 군집을 합쳐가며 점점 큰 군집을 형성한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:32.03125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAAAzElEQVQY03WQ627DIAyF8/7vOCnRIG3DQrj4wjXUi6r9aDUEkg3Hx/6YxhilVtkStN7H/+s8z7d0KqUsy6K19t5vxqSU/p4RsbUmIrmU4NNuqrWu601pHUKIMZqfXdTucefgU8nbfnMBTIBMZFe9G4NMUpJTPpybRGqPQynlnJPmiHT2nlZVzF0owDtxjIidmeavHKMNhggBMOX821mmnef5W6nd2nrBjwsPIfrtAUiH92/Ar7HlELPMDACJ+cV2SVlQry8Ux97b+Kh/Aj+AXGy0rDATAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;louvain&quot; title=&quot;louvain&quot; src=&quot;/static/8191e7a10ee757bea3e0af66e9146836/2bef9/louvain.png&quot; srcSet=&quot;/static/8191e7a10ee757bea3e0af66e9146836/6f3f2/louvain.png 256w,/static/8191e7a10ee757bea3e0af66e9146836/01e7c/louvain.png 512w,/static/8191e7a10ee757bea3e0af66e9146836/2bef9/louvain.png 1024w,/static/8191e7a10ee757bea3e0af66e9146836/ddc6c/louvain.png 1534w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ol&gt;&lt;li&gt;개별 정점으로 구성된 크기 1의 군집들로부터 시작한다.&lt;/li&gt;&lt;li&gt;각 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 기존 혹은 새로운 군집으로 이동한다. 이 때, 군집성이 최대화되도록 군집을 결정한다.&lt;/li&gt;&lt;li&gt;더 이상 군집성이 증가하지 않을 때까지 2를 반복한다.&lt;/li&gt;&lt;li&gt;각 군집을 하나의 정점으로 하는 군집 레벨의 그래프를 얻은 뒤 3을 수행한다.&lt;/li&gt;&lt;li&gt;한개의 정점이 남을 때까지 4를 반복한다.&lt;/li&gt;&lt;/ol&gt;&lt;h2 id=&quot;중첩-군집-구조&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A4%91%EC%B2%A9-%EA%B5%B0%EC%A7%91-%EA%B5%AC%EC%A1%B0&quot; aria-label=&quot;중첩 군집 구조 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;중첩 군집 구조&lt;/h2&gt;&lt;p&gt;위의 Girvan-Newman 알고리즘, Louvain 알고리즘 등은 군집간의 중첩이 없다고 가정하고 있는데, 실제 그래프의 군집들은 중첩되어있는 경우가 많다. 이를 해결하기 위해 &lt;strong&gt;&lt;code&gt;중첩 군집 모형&lt;/code&gt;&lt;/strong&gt;을 따로 정의한다.&lt;/p&gt;&lt;p&gt;아래와 같은 중첩 군집 모형을 가정한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;각 정점은 여러 개의 군집에 속할 수 있다.&lt;/li&gt;&lt;li&gt;각 군집 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대하여, 같은 군집에 속하는 두 정점은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P_A&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;A&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 확률로 간선으로 직접 연결된다.&lt;/li&gt;&lt;li&gt;두 정점이 여러 군집에 동시에 속할 경우 간선 연결 확률은 독립적이다. 예를 들어, 군집 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;B&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 두 정점이 동시에 속할경우, 두 정점이 간선으로 연결될 확률은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;1-(1-P_A)(1-P_B)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;A&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다.&lt;/li&gt;&lt;li&gt;어느 군집에도 함께 속하지 않는 두 정점은 낮은 확률 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;ϵ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\epsilon&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;ϵ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;으로 직접 연결된다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:26.5625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAFCAIAAADKYVtkAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA/klEQVQY0y2O7U6DMBiFuf+78AaM2X+XGLO4qNEqA2VhjCEdlA5aSGlLP62Z58/75LznJCdyzkFYM8a894hMD4cmwDrdYkYCaKONNZQT56wythnGkRDa4ao7h1dkrUUIUTqG6A/Ct08fAW5e7/Z96Z1NYVJ2RdZ8jYLinjyCXbzZbFerNM/2VR5prXCHUYu8d3Jmks/t1F/M1Ip+1txao7QaWK+sKWAb58cMgOf1PciTuoVRmF2dym4g2RmnEB+Lsv58g0mcVdk7/M7RKQT8Vc7hkYXZE74ME50lj4IphJDLwo2dJyFfcgsOBhRyV3IlhJL/Zef/zqJ9Aj3lV+cXYvEVY4uUtXUAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;nested-communities&quot; title=&quot;nested-communities&quot; src=&quot;/static/6c82c3b0f64aec58c61b9dde59b1a4a8/2bef9/nested-communities.png&quot; srcSet=&quot;/static/6c82c3b0f64aec58c61b9dde59b1a4a8/6f3f2/nested-communities.png 256w,/static/6c82c3b0f64aec58c61b9dde59b1a4a8/01e7c/nested-communities.png 512w,/static/6c82c3b0f64aec58c61b9dde59b1a4a8/2bef9/nested-communities.png 1024w,/static/6c82c3b0f64aec58c61b9dde59b1a4a8/71c1d/nested-communities.png 1536w,/static/6c82c3b0f64aec58c61b9dde59b1a4a8/a878e/nested-communities.png 2048w,/static/6c82c3b0f64aec58c61b9dde59b1a4a8/e2cbf/nested-communities.png 2138w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;중첩 군집 모형(각 정점들이 어떤 확률로 연결되어있는지)이 주어지면, 해당 그래프의 확률을 계산할 수 있다.&lt;/p&gt;&lt;p&gt;그래프의 확률은 다음 확률들의 곱이다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;그래프의 각 간선이 두 정점이 (모형에 의해) 직접 연결될 확률&lt;/li&gt;&lt;li&gt;그래프에서 직접 연결되지 않은 각 정점 쌍이 (모형에 의해)&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;그러나 현실에서, 대부분 중첩 그래프는 있지만, 중첩 군집 모형은 주어지지 않는 경우가 많다. &lt;strong&gt;&lt;code&gt;중첩 군집 탐색&lt;/code&gt;&lt;/strong&gt;은 역으로 &lt;strong&gt;주어진 그래프의 확률을 최대화하는 중첩 군집 모형을 찾는 과정&lt;/strong&gt;이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;통계 용어로는 &lt;strong&gt;&lt;code&gt;최우도 추정치(Maximum Likelihood Estimate)&lt;/code&gt;&lt;/strong&gt;를 찾는 과정이다.&lt;/li&gt;&lt;li&gt;이 때, 각 정점의 특정 군집 소속여부가 이산적(discrete)으로 결정되기 때문이다. 따라서 연속적인 값을 최적화하는데 사용하는 일반적인 경사하강법 등을 사용할 수 없다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;따라서, 중첩 군집 탐색을 용이하게 하기 위해 &lt;strong&gt;&lt;code&gt;완화된 중첩 군집 모형&lt;/code&gt;&lt;/strong&gt;을 사용한다. 이 모형에서는 각 정점이 각 군집에 속해 있는 정도를 0(속하지 않음)과 1(속함)으로 표시하는 것이 아니라, 실숫값으로 표현한다. 중간 상태를 표현할 수 있게 된 것이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;모형의 매개변수(간선 값)들이 실수 값(연속된 값)을 가지므로, 익숙한 최적화 도구(경사하강법 등)을 수행하여 탐색할 수 있다.&lt;/li&gt;&lt;/ul&gt;</content:encoded></item><item><title><![CDATA[Graph 03 - 전파 모형]]></title><description><![CDATA[검색 엔진에서는 그래프를 어떻게 활용할까 by 신기정 교수님, BoostCamp AI Tech 5주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/32_cascade_model/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/32_cascade_model/</guid><pubDate>Tue, 23 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;전파-모델&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A0%84%ED%8C%8C-%EB%AA%A8%EB%8D%B8&quot; aria-label=&quot;전파 모델 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;전파 모델&lt;/h1&gt;&lt;h3 id=&quot;그래프를-통한-정보-전파&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B7%B8%EB%9E%98%ED%94%84%EB%A5%BC-%ED%86%B5%ED%95%9C-%EC%A0%95%EB%B3%B4-%EC%A0%84%ED%8C%8C&quot; aria-label=&quot;그래프를 통한 정보 전파 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;그래프를 통한 정보 전파&lt;/h3&gt;&lt;p&gt;온라인 소셜네트워크를 통해 전파되는 정보가 대표적인 예시이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;트위터, 페이스북등을 통한 뉴스 전파&lt;/li&gt;&lt;li&gt;아이스버킷 챌린지, 펭귄 문제 등&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;또는 컴퓨터 네트워크의 고장 전파 상황도 예로 들 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;특정 장비의 고장은 다른 장비의 과부하로 이어져, 고장이 전파되는 양상을 띈다.&lt;/li&gt;&lt;li&gt;이는 파워그리드에서 정전이 퍼져나가는 현상과 유사하다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;사회에서의 질병 전파도 마찬가지이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;코로나 19, 메르스, 사스&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;의사결정-기반의-전파-모형&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9D%98%EC%82%AC%EA%B2%B0%EC%A0%95-%EA%B8%B0%EB%B0%98%EC%9D%98-%EC%A0%84%ED%8C%8C-%EB%AA%A8%ED%98%95&quot; aria-label=&quot;의사결정 기반의 전파 모형 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;의사결정 기반의 전파 모형&lt;/h2&gt;&lt;h3 id=&quot;언제-사용하는가&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%96%B8%EC%A0%9C-%EC%82%AC%EC%9A%A9%ED%95%98%EB%8A%94%EA%B0%80&quot; aria-label=&quot;언제 사용하는가 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;언제 사용하는가?&lt;/h3&gt;&lt;p&gt;주변 사람의 의사결정이 본인의 의사결정에 영향을 미친다. 주변사람들의 의사결정을 고려하여 각자 의사결정을 내리는 경우 &lt;strong&gt;&lt;code&gt;의사결정 기반의 전파 모형&lt;/code&gt;&lt;/strong&gt;을 사용한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;친구들이 모두 페이스북 메신저를 사용하면, 카카오톡 대신 페이스북 메신저를 사용하게 될 것이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;의사결정 기반의 전파 모형 중 가장 간단한 것으로 &lt;strong&gt;선형 임계치 모형(Linear Threshold Model)&lt;/strong&gt;을 살펴보자.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;선형-임계치-모형&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%84%A0%ED%98%95-%EC%9E%84%EA%B3%84%EC%B9%98-%EB%AA%A8%ED%98%95&quot; aria-label=&quot;선형 임계치 모형 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;선형 임계치 모형&lt;/h3&gt;&lt;p&gt;친구 관계의 두 사람 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 있다고 할 때, 두 사람이 호환되지 않는 기술 A와 B 중 하나를 선택하여 사용하는 경우를 생각해보자.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;두 사람 모두 A를 사용하면, 행복이 a만큼 증가한다.&lt;/li&gt;&lt;li&gt;두 사람 모두 B를 사용하면, 행복이 b만큼 증가한다.&lt;/li&gt;&lt;li&gt;두 사람이 각각 다른 기술을 사용하면, 행복이 증가하지 않는다(서로 교류할 수 없기 때문에)&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;이를 확대시켜 사회에서 적용시켜보면, 각자는 본인과 연결된 수많은 vertex가 어떤 기술을 사용하는지 고려하여 결정을 내리게 된다.&lt;/p&gt;&lt;p&gt;만약 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 주위에 A를 사용하는 사람이 5명이고, B를 사용하는 사람이 8명이라고 생각해보자. &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 선택에 따라 행복은 5a가 될수도, 8b가 될수도 있다. 각자는 행복을 최대화하는 선택을 한다는 가정하에, 5a와 8b중 더 큰것을 고를 것이다.&lt;/p&gt;&lt;p&gt;이를 일반화하여, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 비율의 이웃이 A를 선택했을 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;1-p&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 비율의 이웃이 B를 선택했다고 가정하자. 이 때 A를 선택할 확률은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;ap&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; &amp;gt; &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;b(1-p)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다.&lt;/p&gt;&lt;p&gt;정리하면, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo&gt;&amp;gt;&lt;/mo&gt;&lt;mfrac&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p&amp;gt;\frac{b}{a+b}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7335400000000001em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;&amp;gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.283439em;vertical-align:-0.403331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8801079999999999em&quot;&gt;&lt;span style=&quot;top:-2.655em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.394em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.403331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;일때 A를 선택하며, 이때 편의상 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{b}{a+b}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.283439em;vertical-align:-0.403331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8801079999999999em&quot;&gt;&lt;span style=&quot;top:-2.655em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.394em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.403331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 임계치 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;q&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 한다. 이 모형을 &lt;strong&gt;&lt;code&gt;선형 임계치 모형&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;각 정점은 이웃 중 A를 선택한 비율이 임계치 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;q&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 넘을 때에만 A를 선택한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:71.484375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;linear-threshold-model&quot; title=&quot;linear-threshold-model&quot; src=&quot;/static/9c700a7ab407ec19d744645c9a28202a/2bef9/linear-threshold-model.png&quot; srcSet=&quot;/static/9c700a7ab407ec19d744645c9a28202a/6f3f2/linear-threshold-model.png 256w,/static/9c700a7ab407ec19d744645c9a28202a/01e7c/linear-threshold-model.png 512w,/static/9c700a7ab407ec19d744645c9a28202a/2bef9/linear-threshold-model.png 1024w,/static/9c700a7ab407ec19d744645c9a28202a/e515d/linear-threshold-model.png 1430w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;그런데, 주위 사람들이 다 B를 쓰더라도, 이와 관계없이 A를 고수하는 얼리어답터 vertex &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 있다고 생각해보자. 이 두 vertex가 다른 정점들의 선택에 어떤 영향을 미칠까?&lt;/p&gt;&lt;p&gt;임계치가 55%일때, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 선택을 고려하면 정점들의 선택은 다음과 같이 바뀐다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:73.828125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;linear-threshold-model2&quot; title=&quot;linear-threshold-model2&quot; src=&quot;/static/bebcd857f16c4114c2b2ed5c6ccf94fa/2bef9/linear-threshold-model2.png&quot; srcSet=&quot;/static/bebcd857f16c4114c2b2ed5c6ccf94fa/6f3f2/linear-threshold-model2.png 256w,/static/bebcd857f16c4114c2b2ed5c6ccf94fa/01e7c/linear-threshold-model2.png 512w,/static/bebcd857f16c4114c2b2ed5c6ccf94fa/2bef9/linear-threshold-model2.png 1024w,/static/bebcd857f16c4114c2b2ed5c6ccf94fa/cf8e5/linear-threshold-model2.png 1402w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;순식간에 주변 노드들의 영향을 받아 A기술을 선택하는 비율이 늘었다. 이 전파를 끝까지 지속해보면,&lt;/p&gt;&lt;p&gt;&lt;img src=&quot;./img/graph/linear-threshold-model3.png&quot; alt=&quot;linear-threshold-model3&quot;/&gt;&lt;/p&gt;&lt;p&gt;위와 같이 두개의 정점을 제외하고는 모두 A 기술을 쓰는 것으로 판도가 뒤바뀐다. 이 때 남은 두 노드는 임계치를 넘지 않았기 때문에 바뀌지 않는다.&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;확률적-전파-모형&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%99%95%EB%A5%A0%EC%A0%81-%EC%A0%84%ED%8C%8C-%EB%AA%A8%ED%98%95&quot; aria-label=&quot;확률적 전파 모형 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;확률적 전파 모형&lt;/h2&gt;&lt;h3 id=&quot;언제-사용하는가-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%96%B8%EC%A0%9C-%EC%82%AC%EC%9A%A9%ED%95%98%EB%8A%94%EA%B0%80-1&quot; aria-label=&quot;언제 사용하는가 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;언제 사용하는가?&lt;/h3&gt;&lt;p&gt;의사결정 기반의 전파 모형은 전파 과정을 간단하게 표현할 수 있지만, &amp;#x27;의지&amp;#x27;나 &amp;#x27;결정&amp;#x27;이 들어가지 않은 전파에 대해서는 적합하지 않다. 예를 들어, 코로나 19 바이러스에 걸리는 상황을 가정해보자. A기술과 B기술을 선택하는것과 달리, 코로나 19 바이러스는 아무도 걸리기로 &amp;#x27;결정&amp;#x27;하지 않을 것이다.&lt;/p&gt;&lt;p&gt;이처럼 코로나 전파는 의사결정이 아닌 확률적 과정이기 때문에, &lt;strong&gt;&lt;code&gt;확률적 전파 모형&lt;/code&gt;&lt;/strong&gt;을 고려해야 한다.&lt;/p&gt;&lt;p&gt;확률적 전파 모형의 가장 간단한 것으로 &lt;strong&gt;독립 전파 모형(Independent Cascade Model)&lt;/strong&gt;이 있다.&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;독립적-전파-모형&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%8F%85%EB%A6%BD%EC%A0%81-%EC%A0%84%ED%8C%8C-%EB%AA%A8%ED%98%95&quot; aria-label=&quot;독립적 전파 모형 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;독립적 전파 모형&lt;/h3&gt;&lt;p&gt;방향성이 있고 가중치가 있는 그래프를 가정할 때, 각 간선 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(u,v)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 가중치 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_{uv}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 감염되었을 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 감염시킬 확률이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 감염되지 않았다고 가정한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;서로 다른 이웃이 전염되는 확률은 독립적이므로, 최초 감염자들로부터 전파가 늘어남에 따라 전파확률이 기하급수적으로 늘어나게 된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때 첫 감염자를 seed 집합이라고 한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이러한 형태를 가지는 모델을 &lt;strong&gt;&lt;code&gt;독립적 전파 모형&lt;/code&gt;&lt;/strong&gt;이라고 한다. 이 때 감염자의 회복을 가정하는 SIS, SIR등의 다른 전파모형도 있다.&lt;/p&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;바이럴-마케팅과-전파-최대화-문제&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%B0%94%EC%9D%B4%EB%9F%B4-%EB%A7%88%EC%BC%80%ED%8C%85%EA%B3%BC-%EC%A0%84%ED%8C%8C-%EC%B5%9C%EB%8C%80%ED%99%94-%EB%AC%B8%EC%A0%9C&quot; aria-label=&quot;바이럴 마케팅과 전파 최대화 문제 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;바이럴 마케팅과 전파 최대화 문제&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;바이럴 마케팅&lt;/code&gt;&lt;/strong&gt;은 &lt;strong&gt;소비자들로 하여금 긍정적인 입소문을 내도록 하여 제품을 홍보하는 기법&lt;/strong&gt;으로, 홍보의 시작점이 아주 중요하다. &lt;strong&gt;시작점이 어디인지 따라서 입소문의 전파 범위가 영향을 받기 때문&lt;/strong&gt;이다. 이 시작점으로 소셜 인플루언서들이 자주 채택되는데, 이는 인플루언서들이 높은 광고비를 받는 이유 중 하나다.&lt;/p&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;시드-집합의-중요성&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%8B%9C%EB%93%9C-%EC%A7%91%ED%95%A9%EC%9D%98-%EC%A4%91%EC%9A%94%EC%84%B1&quot; aria-label=&quot;시드 집합의 중요성 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;시드 집합의 중요성&lt;/h3&gt;&lt;p&gt;앞서 설명한 모형에서, 시드집합은 전파 크기에 큰 영향을 미쳤다. 그런데 만약 시드 집합이 막다른 정점에 있었다면, 전파가 거의 이루어지지 않게 된다. 따라서 &lt;strong&gt;전파 크기를 최대화시키기 위해서는 이 시드집합을 잘 골라야 한다.&lt;/strong&gt;&lt;/p&gt;&lt;p&gt;이처럼 그래프, 전파 모형, 시드 집합의 크기가 주어졌을 때 전파를 최대화하는 시드집합을 찾는 문제를 &lt;strong&gt;&lt;code&gt;전파 최대화 (Influence Maximization)&lt;/code&gt;&lt;/strong&gt; 문제라고 부른다.&lt;/p&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;전파-최대화-문제&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A0%84%ED%8C%8C-%EC%B5%9C%EB%8C%80%ED%99%94-%EB%AC%B8%EC%A0%9C&quot; aria-label=&quot;전파 최대화 문제 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;전파 최대화 문제&lt;/h3&gt;&lt;p&gt;그래프의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;|V|&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개의 정점이 있을 경우, 시드 집합 크기가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개로 제한되더라도 경우의 수는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;_{|V|}C_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.03853em;vertical-align:-0.3551999999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34480000000000005em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3551999999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 된다. 이는 굉장히 큰 수치이고, 이론적으로 많은 전파 모형에 의해 전파 최대화 문제는 NP-hard임이 증명되었으므로 &lt;strong&gt;최고의 시드를 바로 구하기는 어렵다.&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;정점의 수가 1000개, 시드 집합의 크기가 10개라고 칠 때 경우의 수는 몇 해(垓)를 넘어간다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;대신에, 최고의 시드 집합에 근사하는 &lt;strong&gt;&lt;code&gt;휴리스틱(heuristics)&lt;/code&gt;&lt;/strong&gt;을 사용해 볼 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;휴리스틱이란, &lt;strong&gt;이론적으로는 증명할 수 없지만 실험적으로는 잘 동작하는 간편추론&lt;/strong&gt;을 일컫는다. &amp;#x27;대충 어림짐작&amp;#x27;쯤으로 표현할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-7&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-7&quot; aria-label=&quot; 7 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;전파-최대화-휴리스틱의-종류&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A0%84%ED%8C%8C-%EC%B5%9C%EB%8C%80%ED%99%94-%ED%9C%B4%EB%A6%AC%EC%8A%A4%ED%8B%B1%EC%9D%98-%EC%A2%85%EB%A5%98&quot; aria-label=&quot;전파 최대화 휴리스틱의 종류 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;전파 최대화 휴리스틱의 종류&lt;/h3&gt;&lt;p&gt;대표적 휴리스틱으로 &lt;strong&gt;&lt;code&gt;정점의 중심성(Node Centrality)&lt;/code&gt;&lt;/strong&gt;를 사용한다. 시드 집합의 크기가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개일 때, 정점의 중심성이 높은 순으로 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개의 정점을 선택하는 방법이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;정점의 중심성으로는 [페이지랭크 점수, 연결 중심성, 근접 중심성, 매개 중심성] 등이 있다.&lt;/li&gt;&lt;li&gt;합리적인 방법이지만, 최고의 시드 집합을 찾는다는 보장은 없다. 다만 실험적으로는 좋은 결과를 낸다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;또, &lt;strong&gt;&lt;code&gt;탐욕 알고리즘(Greedy Algorithm)&lt;/code&gt;&lt;/strong&gt; 역시 많이 사용된다. 시드 집합의 원소, 즉 최초 전파자를 한번에 한명씩 선택하며, 매 순간 시뮬레이션 하여 더 많은 전파를 일으킬 수 있는 시드를 찾아 다음 타겟으로 지목한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때, 전파의 크기를 비교하기 위해 시뮬레이션을 반복하여 평균 값을 사용한다.&lt;/li&gt;&lt;li&gt;이렇게 한 타겟씩 뽑는 것을 뽑은 시드 집합이 목표 크기에 도달할 때까지 반복한다.&lt;/li&gt;&lt;li&gt;따라서, 최초 전파자간 조합을 고려하지 않고 근시안적으로 최초 전파자를 선택하는 과정을 반복한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이처럼 탐욕 알고리즘은 일견 현명하지 않은 판단처럼 보일지 몰라도, 독립 전파 모형에 의해 정확도가 일부 보장된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;탐욕 알고리즘으로 찾은 시드 집합에 의한 평균 전파 크기가, 최고(이상적) 시드집합에 의한 평균 전파크기의 최소 0.632배 이상은 된다는 것이 증명되어있다. 즉, &lt;strong&gt;최저성능이 보장&lt;/strong&gt;되어있다.&lt;/li&gt;&lt;/ul&gt;</content:encoded></item><item><title><![CDATA[Graph 02 - 페이지랭크 알고리즘]]></title><description><![CDATA[검색 엔진에서는 그래프를 어떻게 활용할까 by 신기정 교수님, BoostCamp AI Tech 5주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/31_page_rank/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/31_page_rank/</guid><pubDate>Tue, 23 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;pagerank-알고리즘&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#pagerank-%EC%95%8C%EA%B3%A0%EB%A6%AC%EC%A6%98&quot; aria-label=&quot;pagerank 알고리즘 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;PageRank 알고리즘&lt;/h1&gt;&lt;p&gt;웹은 웹페이지와 하이퍼링크로 구성된 방향성 있는 그래프(Digraph)이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;하이퍼링크는 웹페이지에서 나가는 간선에 해당한다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;구글-이전의-검색엔진&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B5%AC%EA%B8%80-%EC%9D%B4%EC%A0%84%EC%9D%98-%EA%B2%80%EC%83%89%EC%97%94%EC%A7%84&quot; aria-label=&quot;구글 이전의 검색엔진 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;구글 이전의 검색엔진&lt;/h3&gt;&lt;p&gt;웹에서 방대한 양의 자료를 어떻게 하면 효율적으로 검색할 수 있을까? 이에 대해 구글 이전의 검색엔진은 다음과 같은 방식들을 시도했다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;웹을 하나의 거대한 디렉토리로 보고, 카테고리별로 정리하기&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;그러나, 웹페이지의 수가 너무 많기 때문에 카테고리의 수와 깊이가 무한정 커질수밖에 없었다.&lt;/li&gt;&lt;li&gt;또, 카테고리의 구분이 모호하여 저장과 검색이 어려웠다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;키워드에 의존한 검색엔진&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;해당 키워드를 (여러 번) 포함한 웹페이지를 반환한다.&lt;/li&gt;&lt;li&gt;그러나, 검색엔진에 노출되기 위해 관련성 없는 단어들을 넣은 악의적인 웹페이지에 취약했다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이러한 문제점들을 해결하기 위해, 구글의 창업자, 당시 스탠포드 박사과정이었던 래리 페이지와 세르게이 브린은 &lt;em&gt;&lt;a href=&quot;http://ilpubs.stanford.edu:8090/422/1/1999-66.pdf&quot;&gt;The PageRank Citation Ranking : Bringing Order to Web&lt;/a&gt;&lt;/em&gt; 라는 논문을 내어 PageRank 알고리즘을 제안한다.&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;페이지-랭크-알고리즘&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%8E%98%EC%9D%B4%EC%A7%80-%EB%9E%AD%ED%81%AC-%EC%95%8C%EA%B3%A0%EB%A6%AC%EC%A6%98&quot; aria-label=&quot;페이지 랭크 알고리즘 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;페이지 랭크 알고리즘&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;페이지랭크(PageRank) 알고리즘&lt;/code&gt;&lt;/strong&gt; 의 여러 정의와 원리에 대해서 알아보자.&lt;/p&gt;&lt;h3 id=&quot;정의---투표-관점&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A0%95%EC%9D%98---%ED%88%AC%ED%91%9C-%EA%B4%80%EC%A0%90&quot; aria-label=&quot;정의   투표 관점 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;정의 - 투표 관점&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;투표&lt;/code&gt;&lt;/strong&gt;를 기반으로 사용자 키워드와 관련성이 높고 신뢰할 수 있는 웹페이지를 찾는다.&lt;/p&gt;&lt;p&gt;이 때, 투표는 &lt;strong&gt;&amp;#x27;하이퍼링크&amp;#x27;&lt;/strong&gt;를 의미한다. 어떤 웹페이지 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 다른 웹페이지 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로의 하이퍼링크를 포함한다면, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 유용하다고 생각했기 때문일 것이다. 따라서, &lt;div&gt;하이퍼링크로 많이 참조된(즉, 들어오는 간선이 많은) 웹페이지일수록 투표를 많이 받은 웹페이지&lt;/div&gt;로 볼 수 있다.&lt;/p&gt;&lt;p&gt;그러나, 들어오는 간선을 세는것만으로 충분할까? 만약 그렇게 한다면, 하이퍼링크를 많이 참조해 둔 웹페이지를 이곳 저곳 만들어 뿌려두는 것으로 악용할 수 있을 것이다. 따라서 이런 경우를 걸러내기 위해, &lt;strong&gt;&amp;#x27;가중 투표&amp;#x27;&lt;/strong&gt;를 실시한다. 관련성이 높고 신뢰할 수 있는 웹사이트의 투표에 더 가중치를 두어 계산한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때 &amp;#x27;관련성&amp;#x27;과 &amp;#x27;신뢰성&amp;#x27;의 개념도 결국 투표로 결정될 수 있으므로, 재귀(recursion)와 연립방정식을 통해 이를 구현한다.&lt;ul&gt;&lt;li&gt;어떤 페이지의 신뢰성(페이지랭크)는 결국 다른 페이지의 신뢰성(페이지랭크)에 따라 결정되는데, 이 과정이 재귀적으로 반복된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:95.3125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;pagerank&quot; title=&quot;pagerank&quot; src=&quot;/static/1e9efcaa1b9696c0f751c239380daeea/2bef9/pagerank.png&quot; srcSet=&quot;/static/1e9efcaa1b9696c0f751c239380daeea/6f3f2/pagerank.png 256w,/static/1e9efcaa1b9696c0f751c239380daeea/01e7c/pagerank.png 512w,/static/1e9efcaa1b9696c0f751c239380daeea/2bef9/pagerank.png 1024w,/static/1e9efcaa1b9696c0f751c239380daeea/7a1fb/pagerank.png 1030w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위의 이미지에서 웹페이지 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;j&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.85396em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 페이지랭크 점수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r_j&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 구하면,  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;/&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;/&lt;/mi&gt;&lt;mn&gt;4&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r_j=r_i/3+r_k/4&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;/&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;3&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;/&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다.&lt;/p&gt;&lt;p&gt;이를 수식으로 일반화하면 다음과 같은 식이 된다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mfrac&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r_j = \sum_{i\in N_{in}(j)}\frac{r_i}{d_{out}(i)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.623565em;vertical-align:-1.516005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.10903em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1075599999999999em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;그런데 이 때, 내가 참조(out)하는 페이지가 나를 참조(in)하고 있을수도 있다. 이 경우 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r_j&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 포함되므로, 결국 해당 페이지의 페이지랭크를 구하려고 하는 데 해당 페이지의 페이지랭크를 알아야하는 방식이 된다.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;정의---random-walk-관점&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A0%95%EC%9D%98---random-walk-%EA%B4%80%EC%A0%90&quot; aria-label=&quot;정의   random walk 관점 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;정의 - Random Walk 관점&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;임의 보행(Random Walk)&lt;/strong&gt;을 통해 웹을 서핑하는 웹 서퍼를 가정한다. 즉, 웹서퍼는 현재 웹페이지의 하이퍼링크 중 하나를 균일한 확률로 클릭한다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;웹 서퍼가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째 방문한 웹페이지가 웹페이지 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;일 확률&lt;/strong&gt;을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_i(t)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 할 때, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(t)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;strong&gt;길이가 웹페이지 수와 같은 확률 분포 벡터&lt;/strong&gt;가 된다. &lt;/p&gt;&lt;ul&gt;&lt;li&gt;즉, (&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 들어오는 간선이 있는) 모든 페이지에서 웹페이지 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를  클릭할 확률 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_i(t)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 모두 모아놓은 벡터이다. 당연히 길이가 웹페이지 수와 같아진다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이를 일반화하면 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;mfrac&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_j(t+1) = \sum_{i\in N_{in}(j)}\textcolor{green}{\frac{\textcolor{blue}{p_i(t)}}{d_{out}(i)}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.9430050000000003em;vertical-align:-1.516005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.10903em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:green&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:green&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;color:green;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:blue&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:blue&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{t+1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69841em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:crimson&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot; style=&quot;color:crimson&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:crimson&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째에 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{j}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.85396em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em;color:crimson&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 방문하기 위해서는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{t}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:crimson&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째에 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{j}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.85396em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em;color:crimson&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 들어오는 이웃(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{i}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:crimson&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)에 있어야 한다. - &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{blue}{p_i(t)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:blue&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:blue&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;페이지 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{i}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:crimson&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서 나가는 간선 중 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{j}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.85396em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em;color:crimson&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 향하는 간선(하이퍼링크)을 클릭할 확률은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;/&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{green}{1/d_{out}(i)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;/&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:green&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:green&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다.&lt;/li&gt;&lt;li&gt;따라서, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{t}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:crimson&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째에 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{i}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:crimson&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 있다가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{t+1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69841em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:crimson&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot; style=&quot;color:crimson&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:crimson&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째에 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{j}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.85396em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em;color:crimson&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 옮겨올 확률은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;/&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{green}{\textcolor{blue}{p_i(t)}/d_{out}(i)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:blue&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:blue&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;/&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:green&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:green&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 된다.&lt;/li&gt;&lt;li&gt;이를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{j}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.85396em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em;color:crimson&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 들어오는 모든 간선(하이퍼링크)에 대해서 계산하여 합치면, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{t+1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69841em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:crimson&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot; style=&quot;color:crimson&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:crimson&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째에 페이지 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;crimson&quot;&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{crimson}{j}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.85396em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em;color:crimson&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 들어올 확률이 된다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 때, 웹 서퍼가 충분히 많이 서핑을 하게 되면(즉, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 무한히 커지면) 확률 분포는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(t)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 수렴하게 된다. 다시말해, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{red}{p(t)=p(t+1)=p}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot; style=&quot;color:red&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot; style=&quot;color:red&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot; style=&quot;color:red&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 성립하게 된다.&lt;/p&gt;&lt;p&gt;수렴한 확률분포 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{red}{p}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 정상 분포(Stationary Distribution)이라고 부르고, 이를 통해 위의 수식을 다음과 같이 바꿀 수 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;⇒&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mfrac&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_j(t+1) = \sum_{i\in N_{in}(j)}\frac{p_i(t)}{d_{out}(i)} \Rarr \textcolor{red}{p_j = \sum_{i\in N_{in}(j)}\frac{p_i}{d_{out}(i)}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.9430050000000003em;vertical-align:-1.516005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.10903em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;⇒&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em;color:red&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot; style=&quot;color:red&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.623565em;vertical-align:-1.516005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:red&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em;color:red&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.10903em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em;color:red&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;color:red&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1075599999999999em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;color:red;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이 수식은 투표 관점에서 정의했던 페이지 랭크 점수와 동일하다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;munder&gt;&lt;munder&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mfrac&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mrow&gt;&lt;mtext&gt;투표관점 페이지랭크점수 &lt;/mtext&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mspace width=&quot;1em&quot;&gt;&lt;/mspace&gt;&lt;mspace width=&quot;1em&quot;&gt;&lt;/mspace&gt;&lt;munder&gt;&lt;munder&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mfrac&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mrow&gt;&lt;mtext&gt;임의보행관점 정상분포 &lt;/mtext&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\underbrace{r_j = \sum_{i\in N_{in}(j)}\frac{r_i}{d_{out}(i)}}_{투표 관점\ 페이지랭크점수\ r} \quad\quad \underbrace{p_j = \sum_{i\in N_{in}(j)}\frac{p_i}{d_{out}(i)}}_{임의보행관점\ 정상분포\ p}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:4.086004000000001em;vertical-align:-2.9784440000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1075599999999999em&quot;&gt;&lt;span style=&quot;top:-0.26522399999999946em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.10756em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;투&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;표&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;관&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;점&lt;/span&gt;&lt;span class=&quot;mspace mtight&quot;&gt;&lt;span class=&quot;mtight&quot;&gt; &lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;페&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;이&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;지&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;랭&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;크&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;점&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;수&lt;/span&gt;&lt;span class=&quot;mspace mtight&quot;&gt;&lt;span class=&quot;mtight&quot;&gt; &lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.10756em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.10756em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1075599999999999em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-0.9435549999999995em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.10756em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.1075599999999994em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.10756em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.10903em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1075599999999999em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.1640050000000004em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.8423360000000004em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:1em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:1em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1075599999999999em&quot;&gt;&lt;span style=&quot;top:-0.26522399999999935em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.10756em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;임&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;의&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;보&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;행&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;관&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;점&lt;/span&gt;&lt;span class=&quot;mspace mtight&quot;&gt;&lt;span class=&quot;mtight&quot;&gt; &lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;정&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;상&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;분&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;포&lt;/span&gt;&lt;span class=&quot;mspace mtight&quot;&gt;&lt;span class=&quot;mtight&quot;&gt; &lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.10756em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.10756em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1075599999999999em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-0.9435549999999995em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.10756em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.1075599999999994em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.10756em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.10903em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1075599999999999em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.1640050000000004em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.9784440000000005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;h2 id=&quot;페이지-랭크의-계산&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%8E%98%EC%9D%B4%EC%A7%80-%EB%9E%AD%ED%81%AC%EC%9D%98-%EA%B3%84%EC%82%B0&quot; aria-label=&quot;페이지 랭크의 계산 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;페이지 랭크의 계산&lt;/h2&gt;&lt;h3 id=&quot;반복곱&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%B0%98%EB%B3%B5%EA%B3%B1&quot; aria-label=&quot;반복곱 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;반복곱&lt;/h3&gt;&lt;p&gt;페이지랭크를 정의대로 계산하기 위해 &lt;strong&gt;&lt;code&gt;반복곱(Power Iteration)&lt;/code&gt;&lt;/strong&gt;을 사용한다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;&lt;p&gt;각 웹페이지 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 페이지랭크 점수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r_j^{(0)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.4577719999999998em;vertical-align:-0.4129719999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4231360000000004em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.4129719999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 동일하게 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;/&lt;/mi&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;#&lt;/mi&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;_&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;1/{\#page\_num}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.06em;vertical-align:-0.31em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;/&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;#&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.02778em&quot;&gt;_&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 초기화한다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;아래 식을 이용하여 각 웹페이지의 페이지랭크 점수를 갱신한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mfrac&gt;&lt;msubsup&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r_j^{(t+1)} = \sum_{i\in N_{in}(j)}\frac{r_i^{(t)}}{d_{out}(i)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.4577719999999998em;vertical-align:-0.412972em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4231360000000004em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.412972em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.237805em;vertical-align:-1.516005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.10903em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7218em&quot;&gt;&lt;span style=&quot;top:-2.3588em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0448em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2748em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0448em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.7218em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0448em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4231360000000004em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.27686399999999994em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;페이지랭크 점수가 수렴(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo&gt;≈&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r^{(t)}\approx r^{(t+1)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8879999999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8879999999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)하였으면 종료하고, 아니면 2를 다시 수행한다.&lt;/p&gt;&lt;/li&gt;&lt;/ol&gt;&lt;h3 id=&quot;문제점&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%AC%B8%EC%A0%9C%EC%A0%90&quot; aria-label=&quot;문제점 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;문제점&lt;/h3&gt;&lt;p&gt;그런데, 이 반복곱이 항상 수렴한다고 보장할 수 있을까?&lt;/p&gt;&lt;ul&gt;&lt;li&gt;수렴하지 않고 진동하는 경우도 있다.&lt;/li&gt;&lt;li&gt;이는 들어오는 간선은 있지만, 나가는 간선은 없는 정점 집합인 &lt;strong&gt;&lt;code&gt;스파이더 트랩(Spider Trap)&lt;/code&gt;&lt;/strong&gt;에 의한 문제이다. (&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mo&gt;⇄&lt;/mo&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;a\rightleftarrows b&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8558em;vertical-align:-0.1808em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel amsrm&quot;&gt;⇄&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 의 무한 반복)&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;또, 수렴한다고 하더라도 &amp;quot;합리적인&amp;quot; 점수로 수렴하는 것을 보장할 수 있을까?&lt;/p&gt;&lt;ul&gt;&lt;li&gt;페이지 랭크의 점수가 0으로 수렴하는 경우도 있다.&lt;/li&gt;&lt;li&gt;들어오는 간선은 있지만, 나가는 간선은 없는 &lt;strong&gt;&lt;code&gt;막다른 정점(Dead End)&lt;/code&gt;&lt;/strong&gt;에 의한 문제이다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;해결책&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%95%B4%EA%B2%B0%EC%B1%85&quot; aria-label=&quot;해결책 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;해결책&lt;/h3&gt;&lt;p&gt;위에서 제기된 문제들을 해결하기 위해, &lt;strong&gt;&lt;code&gt;순간이동(Teleport)&lt;/code&gt;&lt;/strong&gt;를 도입하여 탈출 시나리오를 끼워넣는다.&lt;/p&gt;&lt;p&gt;예를 들어, Random Walk 관점에서는 다음과 같이 행동한다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;현재 웹페이지에 하이퍼링크가 있는지 확인한다.&lt;ol&gt;&lt;li&gt;없다면, 임의의 웹페이지로 순간이동하고, 처음부터 다시 시작한다.&lt;/li&gt;&lt;li&gt;있다면, 앞면이 나올 확률이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;α&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\alpha&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.0037em&quot;&gt;α&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;인 동전을 던진다.&lt;/li&gt;&lt;/ol&gt;&lt;/li&gt;&lt;li&gt;동전이 앞면이라면, 하이퍼링크 중 하나를 Random Walk로 클릭한다.&lt;/li&gt;&lt;li&gt;동전이 뒷면이라면, 임의의 웹페이지로 순간이동한다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;순간이동을 통한 탈출로 spider trap이나 dead end에 갇히는 일이 없어졌다. &lt;/p&gt;&lt;p&gt;이 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;α&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\alpha&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.0037em&quot;&gt;α&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 &lt;strong&gt;&lt;code&gt;감폭비율(Damping Factor)&lt;/code&gt;&lt;/strong&gt;이라고 부른다.&lt;/p&gt;&lt;div&gt;&lt;p&gt;감폭비율은 일반적으로 0.8정도를 사용한다.&lt;/p&gt;&lt;/div&gt;&lt;p&gt;이제, 페이지랭크 점수 계산을 순간이동을 도입시켜 수정해보자.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;각 막다른 정점에서 (자신을 포함한) 모든 다른 정점으로 가는 간선을 추가한다.&lt;/li&gt;&lt;li&gt;아래 수식을 사용하여 반복곱을 수행한다.&lt;/li&gt;&lt;/ol&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;munder&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;α&lt;/mi&gt;&lt;mfrac&gt;&lt;msub&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mrow&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mtext&gt; 하이퍼링크&lt;/mtext&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;munder&gt;&lt;munder&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;α&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mstyle&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mrow&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mtext&gt; 순간이동&lt;/mtext&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r_j = \underbrace{\textcolor{blue}{\sum_{i\in N_{in}(j)}\bigg(\alpha\frac{r_i}{d_{out}(i)}\bigg)}}_{by\ 하이퍼링크} + \underbrace{\textcolor{red}{(1-\alpha)\frac{1}{|V|}}}_{by\ 순간이동}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:4.436221em;vertical-align:-2.9862210000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.4499999999999997em&quot;&gt;&lt;span style=&quot;top:-0.5998869999999993em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.4499999999999997em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mspace mtight&quot;&gt;&lt;span class=&quot;mtight&quot;&gt; &lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;하&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;이&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;퍼&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;링&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;크&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.4499999999999997em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.4499999999999997em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.4499999999999997em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-1.2859949999999998em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.45em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.808995em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:blue&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em;color:blue&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.10903em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot; style=&quot;color:blue&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em;color:blue&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot; style=&quot;color:blue&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;color:blue&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.516005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;delimsizing size3&quot; style=&quot;color:blue&quot;&gt;&lt;span style=&quot;color:blue&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.0037em;color:blue&quot;&gt;α&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1075599999999999em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:blue&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:blue&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;color:blue;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em;color:blue&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;delimsizing size3&quot; style=&quot;color:blue&quot;&gt;&lt;span style=&quot;color:blue&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.1640050000000004em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.9862210000000005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.7276559999999996em;vertical-align:-2.406216em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3214399999999997em&quot;&gt;&lt;span style=&quot;top:-1.0513319999999997em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.32144em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mspace mtight&quot;&gt;&lt;span class=&quot;mtight&quot;&gt; &lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;순&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;간&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;이&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback mtight&quot;&gt;동&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.32144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.32144em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-1.7374399999999999em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.32144em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.32144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.32144em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot; style=&quot;color:red&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.0037em;color:red&quot;&gt;α&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em;color:red&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;color:red;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.584em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.406216em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;&lt;img src=&quot;./img/graph/pagerank2.png&quot; alt=&quot;pagerank2&quot;/&gt;&lt;/p&gt;&lt;p&gt;위의 이미지는 수정된 페이지 랭크의 예시이다. 다음과 같은 특이점을 발견할 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;B는 여러 페이지로부터 투표(참조)를 받았으므로 점수가 높다.&lt;/li&gt;&lt;li&gt;점수가 낮은 페이지(보라색)들은 들어오는 간선이 없다. 그럼에도 불구하고 점수가 존재하는 이유는, 순간이동을 통해 들어올 가능성이 존재하기 때문이다.&lt;/li&gt;&lt;li&gt;C는 들어오는 간선이 하나밖에 없음에도, 신뢰성이 높은 페이지 B가 참조하는 단 하나의 링크이기때문에 점수가 높다.&lt;/li&gt;&lt;/ul&gt;</content:encoded></item><item><title><![CDATA[Graph 01 - 그래프 기초와 패턴]]></title><description><![CDATA[그래프 이론 기초 & 그래프 패턴 by 신기정 교수님, BoostCamp AI Tech 5주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/30_graph_basic/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/30_graph_basic/</guid><pubDate>Mon, 22 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;graph-basic&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#graph-basic&quot; aria-label=&quot;graph basic permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Graph Basic&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;그래프(Graph)&lt;/code&gt;&lt;/strong&gt;는 &lt;code&gt;정점(Vertex)&lt;/code&gt; 집합과, &lt;code&gt;간선(Edge)&lt;/code&gt; 집합으로 이루어진 수학적 구조이다. 네트워크로도 불리며, 정점은 노드(Node)로, 간선은 링크(Link)로도 불린다.&lt;/p&gt;&lt;p&gt;우리 주위에 있는 &lt;strong&gt;&lt;code&gt;복잡계(Complex System)&lt;/code&gt;&lt;/strong&gt;가 가진 공통적인 특성은 &lt;strong&gt;구성요소간의 복잡한 상호작용&lt;/strong&gt;이다. 그래프는 이 복잡계의 상호작용을 효과적으로 표현하고, 복잡계를 분석하기 위한 언어이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;뇌, 지식 그래프, 화학 분자, 단백질 상호작용, 세포간 유사도 그래프, 이미지 분해...&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;그래프를-활용한-ai-task&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B7%B8%EB%9E%98%ED%94%84%EB%A5%BC-%ED%99%9C%EC%9A%A9%ED%95%9C-ai-task&quot; aria-label=&quot;그래프를 활용한 ai task permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;그래프를 활용한 AI Task&lt;/h2&gt;&lt;h3 id=&quot;정점-분류node-classification-문제&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A0%95%EC%A0%90-%EB%B6%84%EB%A5%98node-classification-%EB%AC%B8%EC%A0%9C&quot; aria-label=&quot;정점 분류node classification 문제 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;정점 분류(Node Classification) 문제&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;트위터의 공유(Retweet) 관계를 분석하여 사용자의 정치적 성향 파악하기&lt;/li&gt;&lt;li&gt;단백질의 상호작용을 표현하여 단백질의 역할 알아내기&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;연결-예측linked-prediction-문제&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%97%B0%EA%B2%B0-%EC%98%88%EC%B8%A1linked-prediction-%EB%AC%B8%EC%A0%9C&quot; aria-label=&quot;연결 예측linked prediction 문제 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;연결 예측(Linked Prediction) 문제&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;거시적 - 페이스북 소셜네트워크는 진화 방향 예측하기&lt;/li&gt;&lt;li&gt;미시적 : 추천문제 - 고객에게 필요한 물건을 추천하고, 어떤 물건을 구매했을 때 만족도가 높을지 예측하기&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;군집-분석community-detection-문제&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B5%B0%EC%A7%91-%EB%B6%84%EC%84%9Dcommunity-detection-%EB%AC%B8%EC%A0%9C&quot; aria-label=&quot;군집 분석community detection 문제 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;군집 분석(Community Detection) 문제&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;연결관계로부터 사회적 무리(Social Circle) 찾아내기&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;랭킹ranking-및-정보-검색information-retrieval-문제&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%9E%AD%ED%82%B9ranking-%EB%B0%8F-%EC%A0%95%EB%B3%B4-%EA%B2%80%EC%83%89information-retrieval-%EB%AC%B8%EC%A0%9C&quot; aria-label=&quot;랭킹ranking 및 정보 검색information retrieval 문제 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;랭킹(Ranking) 및 정보 검색(Information Retrieval) 문제&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;웹(Web)을 그래프로 보고, 중요한(필요한) 웹페이지를 찾아내기&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;정보-전파information-cascading-및-바이럴-마케팅viral-marketing-문제&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A0%95%EB%B3%B4-%EC%A0%84%ED%8C%8Cinformation-cascading-%EB%B0%8F-%EB%B0%94%EC%9D%B4%EB%9F%B4-%EB%A7%88%EC%BC%80%ED%8C%85viral-marketing-%EB%AC%B8%EC%A0%9C&quot; aria-label=&quot;정보 전파information cascading 및 바이럴 마케팅viral marketing 문제 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;정보 전파(Information Cascading) 및 바이럴 마케팅(Viral Marketing) 문제&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;네트워크를 통한 정보전달을 최대화하기&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;그래프-관련-필수-기초-개념&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B7%B8%EB%9E%98%ED%94%84-%EA%B4%80%EB%A0%A8-%ED%95%84%EC%88%98-%EA%B8%B0%EC%B4%88-%EA%B0%9C%EB%85%90&quot; aria-label=&quot;그래프 관련 필수 기초 개념 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;그래프 관련 필수 기초 개념&lt;/h2&gt;&lt;h3 id=&quot;유형-및-분류&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9C%A0%ED%98%95-%EB%B0%8F-%EB%B6%84%EB%A5%98&quot; aria-label=&quot;유형 및 분류 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;유형 및 분류&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;방향의 유무에 따라&lt;ul&gt;&lt;li&gt;&lt;code&gt;방향이 없는 그래프(Undirected Graph)&lt;/code&gt;&lt;ul&gt;&lt;li&gt;간선에 방향이 없음&lt;/li&gt;&lt;li&gt;ex) 협업관계 그래프, 페이스북 친구 그래프&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;방향이 있는 그래프(Directed Graph)&lt;/code&gt;&lt;ul&gt;&lt;li&gt;간선에 방향이 있음&lt;/li&gt;&lt;li&gt;ex) 인용 그래프, 트위터 팔로우 그래프&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;가중치의 유무에 따라&lt;ul&gt;&lt;li&gt;&lt;code&gt;가중치가 없는 그래프(Unweighted Graph)&lt;/code&gt;&lt;ul&gt;&lt;li&gt;간선에 가중치가 없음&lt;/li&gt;&lt;li&gt;ex) 웹 그래프, 페이스북 친구 그래프&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;가중치가 있는 그래프(Weighted Graph)&lt;/code&gt;&lt;ul&gt;&lt;li&gt;간선에 가중치가 있음&lt;/li&gt;&lt;li&gt;ex) 전화(횟수) 그래프, 유사도 그래프&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;정점의 종류 개수에 따라&lt;ul&gt;&lt;li&gt;&lt;code&gt;동종 그래프(Unpartite Graph)&lt;/code&gt;&lt;ul&gt;&lt;li&gt;단일 종류의 정점을 가짐&lt;/li&gt;&lt;li&gt;ex) 웹 그래프, 페이스북 친구 그래프&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;이종 그래프(Bipartite Graph)&lt;/code&gt;&lt;ul&gt;&lt;li&gt;두 종류의 정점을 가짐 - 다른 종류의 정점 사이에만 간선이 연결됨&lt;/li&gt;&lt;li&gt;ex) 전자 상거래 구매내역(사용자-상품), 영화 출연 그래프(배우-영화)&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;표현-방식&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%91%9C%ED%98%84-%EB%B0%A9%EC%8B%9D&quot; aria-label=&quot;표현 방식 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;표현 방식&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;일반적으로 정점들의 집합을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;V&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, 간선들의 집합을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;E&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, 그래프를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;G = (V,E)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 적는다.&lt;/li&gt;&lt;li&gt;이웃(Neighbor)는 해당 정점과 연결된 다른 정점을 의미한다. 정점 v의 이웃들의 집합을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N(v)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 혹은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N_v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10903em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 표시한다.&lt;ul&gt;&lt;li&gt;방향성이 있는 그래프에서는 나가는 이웃과 들어오는 이웃을 구분한다.&lt;/li&gt;&lt;li&gt;나가는 이웃들의 집합을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N_{out}(v)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10903em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, 들어오는 이웃들의 집합을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N_{in}(v)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10903em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 적는다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;간선 리스트(Edge List) - 그래프를 간선들의 리스트로 저장한다.&lt;ul&gt;&lt;li&gt;간선은 해당 간선이 연결하는 두 정점들의 순서쌍(Pair)로 저장된다.&lt;ul&gt;&lt;li&gt;ex) (1,2), (5,7)&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;방향성이 있는 간선의 경우 (출발점, 도착점) 순서로 저장된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;인접 리스트(Adjacent List) - 각 정점들의 이웃들을 리스트로 저장한다.&lt;ul&gt;&lt;li&gt;방향성이 없는 경우&lt;ul&gt;&lt;li&gt;ex) {1: 2,4,5 }&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;방향성이 있는 경우&lt;ul&gt;&lt;li&gt;ex) 나가는 이웃 {1 : 2,4}, 들어오는 이웃 {1: 4, 2 : 1, 4 : 1}&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;인접 행렬(Adjacent Matrix) - 방향성이 없는 경우&lt;ul&gt;&lt;li&gt;정점 수 X 정점 수 크기의 행렬&lt;/li&gt;&lt;li&gt;방향성이 있는 경우&lt;ul&gt;&lt;li&gt;정점 i와 j 사이에 간선이 있는 경우, 행렬 i행 j열(그리고 j행 i열)의 원소가 1, 그렇지 않으면 0&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;방향성이 없는 경우&lt;ul&gt;&lt;li&gt;정점 i에서 j로의 간선이 있는 경우, 행렬의 i행 j열 원소가 1, 그렇지 않으면 0&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;관련-라이브러리&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B4%80%EB%A0%A8-%EB%9D%BC%EC%9D%B4%EB%B8%8C%EB%9F%AC%EB%A6%AC&quot; aria-label=&quot;관련 라이브러리 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;관련 라이브러리&lt;/h3&gt;&lt;p&gt;&lt;code&gt;NetworkX&lt;/code&gt; - 비교적 속도가 느리지만, 사용이 편리하다.&lt;/p&gt;&lt;p&gt;&lt;code&gt;Snap.py&lt;/code&gt; - 비교적 속도가 빠르지만, 사용이 어렵다.&lt;/p&gt;&lt;h1 id=&quot;graph-pattern&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#graph-pattern&quot; aria-label=&quot;graph pattern permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Graph Pattern&lt;/h1&gt;&lt;h2 id=&quot;실제-그래프와-랜덤-그래프&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%8B%A4%EC%A0%9C-%EA%B7%B8%EB%9E%98%ED%94%84%EC%99%80-%EB%9E%9C%EB%8D%A4-%EA%B7%B8%EB%9E%98%ED%94%84&quot; aria-label=&quot;실제 그래프와 랜덤 그래프 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;실제 그래프와 랜덤 그래프&lt;/h2&gt;&lt;h3 id=&quot;실제-그래프&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%8B%A4%EC%A0%9C-%EA%B7%B8%EB%9E%98%ED%94%84&quot; aria-label=&quot;실제 그래프 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;실제 그래프&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;실제 그래프(Real Graph)&lt;/code&gt;&lt;/strong&gt;란, 다양한 복잡계로부터 얻어진 그래프를 의미한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;소셜 네트워크, 전자상거래 구매 내역, 인터넷 ,웹, 뇌 ...&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;랜덤-그래프&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%9E%9C%EB%8D%A4-%EA%B7%B8%EB%9E%98%ED%94%84&quot; aria-label=&quot;랜덤 그래프 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;랜덤 그래프&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;랜덤 그래프(Random Graph)&lt;/code&gt;&lt;/strong&gt;란, 확률적 과정을 통해 생성한 그래프를 의미한다. 실제 그래프와 비교하기 위한 대상으로서 사용된다.&lt;/p&gt;&lt;p&gt;&lt;code&gt;에르되스-레니 랜덤 그래프&lt;/code&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;가장 간단한 형태의 랜덤그래프로, 임의의 두 정점 사이에 간선이 존재하는지 여부는 동일한 확률분포에 의해 결정된다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;G(n,p)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개의 정점을 가진다.&lt;/li&gt;&lt;li&gt;임의의 두 정점 사이에 간선이 존재할 확률은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다.&lt;/li&gt;&lt;li&gt;정점 간의 연결은 서로 독립적(Independent)이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-7&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-7&quot; aria-label=&quot; 7 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;경로와-거리&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B2%BD%EB%A1%9C%EC%99%80-%EA%B1%B0%EB%A6%AC&quot; aria-label=&quot;경로와 거리 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;경로와 거리&lt;/h2&gt;&lt;h3 id=&quot;개념&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B0%9C%EB%85%90&quot; aria-label=&quot;개념 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;개념&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 사이의 &lt;strong&gt;&lt;code&gt;경로(Path)&lt;/code&gt;&lt;/strong&gt;는 아래 조건을 만족하는 정점들의 순열(Sequence)이다.&lt;ol&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서 시작하여 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서 끝난다.&lt;/li&gt;&lt;li&gt;순열에서 연속된 정점은 간선으로 연결되어 있어야 한다.&lt;/li&gt;&lt;/ol&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;경로의 길이&lt;/code&gt;는 해당 경로 상에 놓이는 간선의 수로 정의된다.&lt;/li&gt;&lt;li&gt;정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 사이의 &lt;code&gt;거리(Distance)&lt;/code&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;u&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 사이 최단경로의 길이이다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;지름(Diameter)&lt;/code&gt;은 정점 간 거리의 최댓값이다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-8&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-8&quot; aria-label=&quot; 8 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;작은-세상-효과&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9E%91%EC%9D%80-%EC%84%B8%EC%83%81-%ED%9A%A8%EA%B3%BC&quot; aria-label=&quot;작은 세상 효과 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;작은 세상 효과&lt;/h3&gt;&lt;p&gt;스탠리 밀그램의 &lt;strong&gt;여섯 단계 분리(Six Degrees of Separation) 실험&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;임의의 두 사람 간의 관계는 평균적으로 6단계로 표현되었다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;사람들이 서로 소수의 공통된 지인을 통해 연결되어 있다는 사실, 즉, 임의의 두 정점 사이의 거리가 작다는 사실을 &lt;code&gt;작은 세상 효과(Small-world Effect)&lt;/code&gt;라고 부른다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;중복이 없다고 가정하였을 때, 모든 사람에게 100명의 지인이 있다면, 다섯 단계를 거치면 100억(=100^5)명까지도 연결될 수 있다.&lt;/li&gt;&lt;li&gt;물론 실제 세상에서는 중복이 있으므로 이보다 적을 것이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;작은 세상 효과는 높은 확률로 랜덤그래프에도 존재하는데, 한 정점 당 이웃의 수가 늘어날수록, 적은 거리로도 연결되는 정점의 수가 기하급수적으로 증가하기 때문이다.&lt;/p&gt;&lt;p&gt;그렇다고 모든 그래프에서 작은 세상 효과가 존재하지는 않는다. 체인(Chain), 사이클(Cycle), 격자(Grid) 그래프에서는 작은 세상효과가 없다.&lt;/p&gt;&lt;h1 id=&quot;-9&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-9&quot; aria-label=&quot; 9 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;연결성&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%97%B0%EA%B2%B0%EC%84%B1&quot; aria-label=&quot;연결성 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;연결성&lt;/h2&gt;&lt;h3 id=&quot;개념-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B0%9C%EB%85%90-1&quot; aria-label=&quot;개념 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;개념&lt;/h3&gt;&lt;p&gt;정점의 &lt;strong&gt;&lt;code&gt;연결성(Degree)&lt;/code&gt;&lt;/strong&gt;는 그 정점과 연결된 간선의 수를 의미한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 연결성은 해당 정점의 이웃들의 수와 같다.&lt;/li&gt;&lt;li&gt;보통 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 연결성을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d(v)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d_v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, 혹은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;|N(v)|&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 적는다.&lt;/li&gt;&lt;li&gt;나가는 연결성은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d_{out}(v)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;|N_{out}(v)|&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10903em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 표시하며, 들어오는 연결성은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d_{in}(v)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;|N_{in}(v)|&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10903em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 표시한다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-10&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-10&quot; aria-label=&quot; 10 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;두터운-꼬리-분포&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%91%90%ED%84%B0%EC%9A%B4-%EA%BC%AC%EB%A6%AC-%EB%B6%84%ED%8F%AC&quot; aria-label=&quot;두터운 꼬리 분포 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;두터운 꼬리 분포&lt;/h3&gt;&lt;p&gt;실제 그래프의 연결성 분포는 &lt;code&gt;두터운 꼬리(Heavy Tail)&lt;/code&gt;를 갖는다. 즉, 연결성이 매우 높은 &lt;code&gt;허브(Hub)&lt;/code&gt; 정점이 존재한다.&lt;/p&gt;&lt;p&gt;그러나, 랜덤 그래프의 연결성 분포는 높은 확률로 정규분포와 유사하다. 즉, 허브 정점이 존재할 가능성이 0에 가깝다.&lt;/p&gt;&lt;h1 id=&quot;-11&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-11&quot; aria-label=&quot; 11 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;연결-요소&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%97%B0%EA%B2%B0-%EC%9A%94%EC%86%8C&quot; aria-label=&quot;연결 요소 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;연결 요소&lt;/h2&gt;&lt;h3 id=&quot;개념-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B0%9C%EB%85%90-2&quot; aria-label=&quot;개념 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;개념&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;연결 요소(Connected Component)&lt;/code&gt;&lt;/strong&gt;는 다음 조건들을 만족하는 정점들의 집합을 일컫는다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;연결 요소에 속하는 정점들은 경로로 연결될 수 있다.&lt;/li&gt;&lt;li&gt;1의 조건을 만족하면서 정점을 추가할 수 없다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;쉽게 말하면, 뚝 뚝 떨어진 하나의 그래프 덩어리라고 볼 수 있다.&lt;/p&gt;&lt;h1 id=&quot;-12&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-12&quot; aria-label=&quot; 12 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;거대-연결-요소&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B1%B0%EB%8C%80-%EC%97%B0%EA%B2%B0-%EC%9A%94%EC%86%8C&quot; aria-label=&quot;거대 연결 요소 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;거대 연결 요소&lt;/h3&gt;&lt;p&gt;실제 그래프에는 &lt;code&gt;거대 연결 요소(Giant Connected Component)&lt;/code&gt;가 존재한다. 거대 연결 요소는 대다수의 정점을 포함한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;MSN 메신저에서는 99.9%의 정점이 하나의 거대 연결 요소에 포함된다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;랜덤 그래프에도 높은 확률로 거대 연결 요소가 존재한다. 단, 정점들의 평균 연결성이 1보다 충분히 커야한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;자세한 이유는 Random Graph Theory를 참고한다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-13&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-13&quot; aria-label=&quot; 13 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;군집-구조&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B5%B0%EC%A7%91-%EA%B5%AC%EC%A1%B0&quot; aria-label=&quot;군집 구조 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;군집 구조&lt;/h2&gt;&lt;h3 id=&quot;개념-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B0%9C%EB%85%90-3&quot; aria-label=&quot;개념 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;개념&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;군집(Community)&lt;/code&gt;&lt;/strong&gt;란 다음 조건들을 만족하는 정점들의 집합이다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;집합에 속하는 정점 사이에 많은 간선이 존재한다.&lt;/li&gt;&lt;li&gt;집합에 속하는 정점과 그렇지 않은 정점 사이에는 적은 수의 간선이 존재한다.&lt;/li&gt;&lt;/ol&gt;&lt;ul&gt;&lt;li&gt;수학적으로 엄밀한 정의는 아니다. &amp;#x27;많다&amp;#x27;와 &amp;#x27;적다&amp;#x27;의 정의가 모호하기 때문인 듯 하다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;code&gt;지역적 군집 계수(Local Clustering Coefficient)&lt;/code&gt;는 한 정점에서 군집의 형성 정도를 측정한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 지역적 군집 계수는 &lt;strong&gt;정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 이웃 쌍 중 간선으로 직접 연결된 것의 비율&lt;/strong&gt;을 의미하며, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 표현한다.&lt;/li&gt;&lt;li&gt;즉, 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 연결된 이웃 정점들이, (&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 연결되는 간선을 제외하고) 서로 간선들로 많이 연결되어 있을 때 지역적 군집 계수가 높다.&lt;/li&gt;&lt;li&gt;반면, 정점 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;만 허브로 존재하고, 나머지 정점들끼리는 연관관계가 없다면 지역적 군집계수가 낮을 것이다.&lt;/li&gt;&lt;li&gt;연결성이 0인 정점에서는 지역적 군집계수가 정의되지 않는다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;전역 군집 계수(Global Clustering Coefficient)는 전체 그래프에서 군집의 형성 정도를 측정한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;그래프 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;G&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 전역 군집 계수는 &lt;strong&gt;각 정점에서의 지역적 군집 계수의 평균&lt;/strong&gt;이다.&lt;ul&gt;&lt;li&gt;단, 지역적 군집계수가 정의되지 않는 정점은 제외한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-14&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-14&quot; aria-label=&quot; 14 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;높은-군집-계수&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%86%92%EC%9D%80-%EA%B5%B0%EC%A7%91-%EA%B3%84%EC%88%98&quot; aria-label=&quot;높은 군집 계수 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;높은 군집 계수&lt;/h3&gt;&lt;p&gt;실제 그래프에서는 군집 계수가 높다. 즉, &lt;strong&gt;많은 군집이 존재&lt;/strong&gt;한다. 이유에는 여러가지가 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;동질성(Homophily)&lt;/code&gt; : 서로 유사한 정점끼리 간선으로 연결될 가능성이 높다.&lt;ul&gt;&lt;li&gt;ex) 같은 동네에 사는 같은 나이의 아이들이 친구가 되는 경우&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;전이성(Transitivity)&lt;/code&gt; : 공통 이웃이 있는 경우, 공통 이웃이 매개 역할을 해줄 수 있다.&lt;ul&gt;&lt;li&gt;ex) 친구를 서로에게 소개해 주는 경우&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;반면, 랜덤그래프에서는 지역적 혹은 전역 군집계수가 높지 않다. &lt;strong&gt;랜덤그래프에서의 간선 연결이 독립적&lt;/strong&gt;이기 때문이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;랜덤그래프 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;G(n,p)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서의 군집 계수는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다.&lt;/li&gt;&lt;/ul&gt;</content:encoded></item><item><title><![CDATA[NLP 08 - NLP에서의 Self-supervised Pretraining Model 살펴보기]]></title><description><![CDATA[Self-supervised Pre-training Models by 주재걸 교수님, BoostCamp AI Tech 4주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/29_self-supervised_learning/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/29_self-supervised_learning/</guid><pubDate>Fri, 19 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;self-supervised-pre-training-models&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#self-supervised-pre-training-models&quot; aria-label=&quot;self supervised pre training models permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Self-supervised Pre-training Models&lt;/h1&gt;&lt;h2 id=&quot;최신-발전-동향&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%B5%9C%EC%8B%A0-%EB%B0%9C%EC%A0%84-%EB%8F%99%ED%96%A5&quot; aria-label=&quot;최신 발전 동향 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;최신 발전 동향&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;Transformer 및 self-attention 모델은 현재 자연어처리 뿐만 아니라 다른 분야에까지 좋은 성능을 내고있다.&lt;/li&gt;&lt;li&gt;최근에는 스택을 깊게 쌓은 형태의 transformer 모델들이 &lt;code&gt;전이 학습&lt;/code&gt; 등을 통하여 성능을 획기적으로 향상시킨 사례가 나오고 있다.&lt;ul&gt;&lt;li&gt;&lt;code&gt;BERT&lt;/code&gt;, &lt;code&gt;GPT-3&lt;/code&gt;, &lt;code&gt;XLNet&lt;/code&gt;, &lt;code&gt;ALBERT&lt;/code&gt;, &lt;code&gt;RoBERTa&lt;/code&gt;, &lt;code&gt;Reformer&lt;/code&gt;, &lt;code&gt;T5&lt;/code&gt;, &lt;code&gt;ELECTRA&lt;/code&gt;...&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;분야도 넓혀가고 있는 중&lt;ul&gt;&lt;li&gt;추천 시스템, 신약 개발, 컴퓨터 비젼...&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;그러나 단어를 하나하나씩 생성한다는 &lt;code&gt;greedy decoding&lt;/code&gt; 형태를 아직 벗어나지 못하고 있는것도 사실이다.&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;gpt-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gpt-1&quot; aria-label=&quot;gpt 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GPT-1&lt;/h2&gt;&lt;p&gt;일론머스크가 참여한 비영리 재단 Open AI에서 개발한 자연어 처리 모델이다.&lt;/p&gt;&lt;p&gt;하나의 task 뿐만 아니라 &lt;strong&gt;자연어 처리와 관련된 여러 task를 모두 커버할 수 있다&lt;/strong&gt;는 것이 특징이다.&lt;/p&gt;&lt;h3 id=&quot;모델-구조와-작동-과정&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%AA%A8%EB%8D%B8-%EA%B5%AC%EC%A1%B0%EC%99%80-%EC%9E%91%EB%8F%99-%EA%B3%BC%EC%A0%95&quot; aria-label=&quot;모델 구조와 작동 과정 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;모델 구조와 작동 과정&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:49.609375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;gpt-1&quot; title=&quot;gpt-1&quot; src=&quot;/static/afa6a164ec0b26c53de909c4e8b6a091/2bef9/gpt-1.png&quot; srcSet=&quot;/static/afa6a164ec0b26c53de909c4e8b6a091/6f3f2/gpt-1.png 256w,/static/afa6a164ec0b26c53de909c4e8b6a091/01e7c/gpt-1.png 512w,/static/afa6a164ec0b26c53de909c4e8b6a091/2bef9/gpt-1.png 1024w,/static/afa6a164ec0b26c53de909c4e8b6a091/71c1d/gpt-1.png 1536w,/static/afa6a164ec0b26c53de909c4e8b6a091/aeac4/gpt-1.png 1827w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;입력 출력 시퀀스가 별도로 있지 않고, &lt;strong&gt;대량의 웹 데이터&lt;/strong&gt;로부터 추출한 문장을 토대로 기존의 &lt;strong&gt;Language Modeling Task 방식&lt;/strong&gt;으로 12개의 self-attention 블록이 학습되는 형식이다.&lt;/p&gt;&lt;p&gt;그러나 단순히 Lauguage Modeling 뿐만 아니라 다른 Task도 다룰 수 있게 하기 위해 새로운 Task Clasiffier 프레임워크를 제안하고 있다.&lt;/p&gt;&lt;p&gt;문장을 넣을 때 기존에 문장 뒤에 넣어주던 &lt;code&gt;&amp;lt;EOS&amp;gt;&lt;/code&gt; 토큰과 조금 다른 &lt;code&gt;&amp;lt;Extract&amp;gt;&lt;/code&gt; 토큰을 넣어인코딩한다. 인코딩 후 Extract 토큰에 해당하는 인코더의 output을 디코더에 input으로 넣어주어 (linear transformation을 거쳐) task를 위한 정보를 파악한다.&lt;/p&gt;&lt;p&gt;예를 들어 이미지에 나온 내포(Entailment)관계에서는, 문장 A(전제,premise)가 참이면 B(가설,hypothesis)도 참인 경우가 있다. 이 경우 두 문장을 하나의 sequence로 만들되, 문장 사이에는 Delimeter(구분자)를 넣어 분리하여 input으로 삼는다.&lt;/p&gt;&lt;p&gt;Extract 토큰은 처음에는 그냥 단순히 문장 마지막에 추가한 토큰이었지만, Self-attention 학습과정 중에 query로 사용되어 학습된다. 결과적으로는 &lt;strong&gt;task에 필요한 정보들을 입력문장으로부터 적절하게 취합/추출할 수 있는 토큰&lt;/strong&gt;이 된다.&lt;/p&gt;&lt;h3 id=&quot;transfer-learning&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#transfer-learning&quot; aria-label=&quot;transfer learning permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Transfer Learning&lt;/h3&gt;&lt;p&gt;대량의 일반적인 task 데이터와 소량의 특정 task 데이터가 있는 경우, 통합적으로 학습한 GPT-1 모델을 &lt;strong&gt;&lt;code&gt;전이학습(Transfer Learning)&lt;/code&gt;&lt;/strong&gt;형태로 활용한다.&lt;/p&gt;&lt;p&gt;위의 아키텍쳐 이미지에서, 기존의 학습모델은 그대로 두고, output 부분에 main task 분류를 위한 추가적인 레이어(Task Classifier)를 하나 덧붙이는 방식으로 새로운 task에 대한 학습을 시켜줄 수 있다.  이 때 새로 들어오는 레이어는 random intialization을 거친 값으로, 빠르게 학습되어야하는 백지 파라미터이다. 따라서 &lt;strong&gt;기학습된 본체의 학습률을 크게 줄이고 추가된 레이어와 함께 학습시킴&lt;/strong&gt;으로써, Task Classifier 레이어를 빠르게 다른 Task에 적용시킬 수 있게 된다.&lt;/p&gt;&lt;h3 id=&quot;self-supervised-learning&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#self-supervised-learning&quot; aria-label=&quot;self supervised learning permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Self-supervised Learning&lt;/h3&gt;&lt;p&gt;Main task에 사용되는 데이터는 별도의 레이블이 필요하다. 그러나 &lt;strong&gt;기존의 pretrain된 main task 데이터는 별도의 레이블이 필요하지 않은 데이터&lt;/strong&gt;이기 때문에, 데이터를 대량으로 확보하고 학습하기 편하다. 이처럼 &lt;strong&gt;레이블이 필요하지 않은 데이터에서 학습하는 것을&lt;/strong&gt; &lt;strong&gt;&lt;code&gt;자기지도학습(Self-supervised Learning)&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;/p&gt;&lt;p&gt;레이블링이 필요하지 않은 대규모 데이터로부터 얻은 지식을, 소량의 데이터가 있는 task에 전이학습 형태로 제공하여 성능을 향상시킬 수 있다.&lt;/p&gt;&lt;h2 id=&quot;bert&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#bert&quot; aria-label=&quot;bert permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;BERT&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Bidectional Trasnformer(BERT)&lt;/code&gt;&lt;/strong&gt; 모델은 현재까지도 가장 널리 쓰이는 pretraining 모델로, GPT-1과 마찬가지로 &lt;strong&gt;Language Modeling 방식으로 학습&lt;/strong&gt;시킨 모델이다.&lt;/p&gt;&lt;p&gt;기존에 다음 단어를 예측하는 등의 간단한 자기지도학습을 &lt;code&gt;Bi-LSTM(ELMo)&lt;/code&gt;으로 하려는 시도가 있었으나, 지금은 BERT로 다 대체되었다.&lt;/p&gt;&lt;h3 id=&quot;masked-language-model&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#masked-language-model&quot; aria-label=&quot;masked language model permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Masked Language Model&lt;/h3&gt;&lt;p&gt;사실 언어는 앞 뒤 문맥을 다 봐야하는 것인데, 기존의 Language Model은 왼쪽(전) 또는 오른쪽(후)만의 정보를 이용해왔다. 이런 동기에서 나온것이 BERT의 pre-training 방식인 &lt;strong&gt;&lt;code&gt;Masked Language Model(MLM)&lt;/code&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;방식은 이렇다. 데이터가 주어지면, 이 문장 데이터 중 일정 비율(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;%)을 &lt;code&gt;[MASK]&lt;/code&gt; 토큰을 이용하여 Mask 처리하여, 나머지 단어들만을 가지고 이를 맞추는 연습을 한다.이 때 &lt;strong&gt;Masking 할 비율 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 얼마이느냐는 Hyperparameter&lt;/strong&gt;이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 너무 낮으면, 학습 시간 대비 훈련량이 적어 효율이 떨어지거나 학습 속도가 느려질 수 있다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 너무 높으면, 문맥을 제대로 파악할 수 없다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;문제는, pre-training 과정에서는 [MASK] 토큰이 어느 정도 비율로 있었는데, &lt;strong&gt;실제 데이터에서는 [MASK] 토큰이 나올 일이 없다&lt;/strong&gt;는 것이다. 따라서 학습과정과 실제 데이터간의 괴리가 발생한다. 이런 차이점이 학습을 방해하거나 전이학습의 효과를 저해한다.&lt;/p&gt;&lt;p&gt;이를 해결하기 위해, 마스킹할 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;%의 단어들 전부가 아니라 어느 정도만 [MASK] 토큰으로 바꾸는 방식을 생각해볼 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;전체 문장이 총 1000개의 단어이고, 이 중 15%(150개)가 모델이 맞추어내야 하는 단어라고 생각하면,&lt;ul&gt;&lt;li&gt;150개 중 120개는 [MASK]로 바꾼다.&lt;/li&gt;&lt;li&gt;150개 중 15개는 랜덤한 단어로 바꾼다.&lt;ul&gt;&lt;li&gt;이 경우 문법적으로나 맥락으로 이 단어가 옳은지 확인해야 하므로, 학습의 난이도가 조금 더 올라간다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;150개 중 나머지 15개는 원본을 유지한다.&lt;ul&gt;&lt;li&gt;이 경우는 모델이 소신있게 &amp;#x27;이 단어가 맞아&amp;#x27;라고 판단한 셈이 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;next-sentence-prediction&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#next-sentence-prediction&quot; aria-label=&quot;next sentence prediction permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Next Sentence Prediction&lt;/h3&gt;&lt;p&gt;BERT가 도입한 또 다른 pre-training 기법으로 &lt;strong&gt;&lt;code&gt;Next Sentence Prediction&lt;/code&gt;&lt;/strong&gt;이 있다.&lt;/p&gt;&lt;p&gt;두 문장을 뽑아 구분자 &lt;code&gt;[SEP]&lt;/code&gt;을 문장끝에 넣은 뒤, 연속적으로 문장들을 잇는다. 그리고 GPT-1에서의 Extraction 토큰 역할을 하는  &lt;code&gt;[CLS]&lt;/code&gt;  토큰을 이은 문장 가장 앞에 넣어 만든 시퀀스를 가지고, 두 문장이 서로 인접한(다음에 오는) 문장인지, 아니면 관계가 없는 문장인지 판단하는 이진분류를 수행한다.&lt;/p&gt;&lt;h3 id=&quot;요약&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9A%94%EC%95%BD&quot; aria-label=&quot;요약 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;요약&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;모델 구조&lt;ul&gt;&lt;li&gt;(L,H,A) = (Self-attention Block, Incoding Vector Dimension ,Attention Head)&lt;/li&gt;&lt;li&gt;BERT BASE : L=12, H=768, A=12&lt;/li&gt;&lt;li&gt;BERT LLARGE : L=24, H=1024, A=16&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Input representation&lt;ul&gt;&lt;li&gt;WordPiece embeddings&lt;ul&gt;&lt;li&gt;입력 시퀀스를 word 별로 하는 것이 아니라 subword 단위로 임베딩하여 인코딩한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Learned positional embedding&lt;ul&gt;&lt;li&gt;단어의 위치정보를 임베딩시에 matrix도 넣어준다.(positional encoding처럼)&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;[CLS] - Clasffication token&lt;/li&gt;&lt;li&gt;Packed Sentence Embedding [SEP]&lt;/li&gt;&lt;li&gt;Segment Embedding&lt;ul&gt;&lt;li&gt;두 문장을 묶어 하나의 세그먼트로 만들 때, 몇번째 문장인지에 해당하는 세그먼트 임베딩을 추가하여 넣어줌으로써 문장을 구별할 수 있게 한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Pre-trained task&lt;ul&gt;&lt;li&gt;Masked Lanugage Model&lt;/li&gt;&lt;li&gt;Next Sentence Prediction&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;fine-tuning-process&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#fine-tuning-process&quot; aria-label=&quot;fine tuning process permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Fine-tuning Process&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:48.828125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;fine-tuning&quot; title=&quot;fine-tuning&quot; src=&quot;/static/a8015c97fd9847cf97b3e89ff3d865ad/2bef9/fine-tuning.png&quot; srcSet=&quot;/static/a8015c97fd9847cf97b3e89ff3d865ad/6f3f2/fine-tuning.png 256w,/static/a8015c97fd9847cf97b3e89ff3d865ad/01e7c/fine-tuning.png 512w,/static/a8015c97fd9847cf97b3e89ff3d865ad/2bef9/fine-tuning.png 1024w,/static/a8015c97fd9847cf97b3e89ff3d865ad/71c1d/fine-tuning.png 1536w,/static/a8015c97fd9847cf97b3e89ff3d865ad/a878e/fine-tuning.png 2048w,/static/a8015c97fd9847cf97b3e89ff3d865ad/0e288/fine-tuning.png 3018w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;기존에 특정 Task를 처리하도록 pre-train된 모델을, 다른 task를 수행할 수 있도록 조정하여 주는 과정을 &lt;strong&gt;&lt;code&gt;미세조정 과정(Fine-tuning Process)&lt;/code&gt;&lt;/strong&gt;이라고 한다. 기존의 모델구조를 거의 바꾸지 않고도 추가적으로 작은 조정만으로 다른 task를 수행할 수 있기 때문이다.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;gpt-vs-bert&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gpt-vs-bert&quot; aria-label=&quot;gpt vs bert permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GPT vs BERT&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;OpenAI GPT&lt;/code&gt;&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;Unidirectional -&lt;/strong&gt; 다음 단어를 예측하는것이 task이기 때문에, 다음 단어에 접근을 허용하면 안된다.&lt;/li&gt;&lt;li&gt;BookCorpus 데이터로 학습(80억개 단어)&lt;/li&gt;&lt;li&gt;배치 사이즈 32,000개 단어&lt;/li&gt;&lt;li&gt;모든 fine-tuning experiments에 대하여 5e-5라는 동일한 학습률 적용&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;BERT&lt;/code&gt;&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;Bidirectional -&lt;/strong&gt; [MASK] 토큰으로 치환된 단어를 예측하기 위하여 앞뒤 문맥을 모두 사용한다.&lt;/li&gt;&lt;li&gt;BookCorpus와 Wikipedia 데이터로 학습(250억개 단어)&lt;/li&gt;&lt;li&gt;[SEP], [CLS], 세그먼트 임베딩&lt;/li&gt;&lt;li&gt;배치 사이즈 128,000개 단어&lt;/li&gt;&lt;li&gt;task에 따라 각기 다른 학습률 적용&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;일반적으로 기존의 모델들에 비해 BERT가 성능이 전반적으로 좋았다(GLUE 데이터 참조)&lt;/p&gt;&lt;h3 id=&quot;기계독해기반-질의응답&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B8%B0%EA%B3%84%EB%8F%85%ED%95%B4%EA%B8%B0%EB%B0%98-%EC%A7%88%EC%9D%98%EC%9D%91%EB%8B%B5&quot; aria-label=&quot;기계독해기반 질의응답 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;기계독해기반 질의응답&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Machine Reading Comprehension(MRC)&lt;/code&gt;&lt;/strong&gt;, 즉 &lt;strong&gt;&lt;code&gt;Question Answering&lt;/code&gt;&lt;/strong&gt;은 질문만 주어지고 답을 예측하는 task가 아니라, 주어진 질문을 &amp;#x27;잘&amp;#x27; 이해하고, 이에 맞는 정답을 예측해내는 task이다.&lt;/p&gt;&lt;p&gt;이 때 기계가 비교적 독해하기 어려운 질문들을 크라우드소싱으로 모아놓은 데이터셋 &lt;strong&gt;&lt;code&gt;SQuAD(Stanford Question Answering Data)&lt;/code&gt;&lt;/strong&gt;가 있다. 출제자가 지정해놓은 정답에 해당하는 문구가 존재하는 문장이 있고, 기계가 독해하여 문맥에 맞는 정답을 추론한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:17.96875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAECAIAAAABPYjBAAAACXBIWXMAABYlAAAWJQFJUiTwAAAAyUlEQVQI1x2OS3ODIACE/f9/Kqdeesg07RgiPhA1BeQdsCJYW5Jvdnb3tlso+ZDCKumM9kY7rZ4lu9Z+seuPFNpaY4wSKoW0b+k37jHGbdtCCAUh3/PMsvB9yilmztnMKGV8Xq2nddV0HSN0ICMYYYkBnJppmgY8UEoL1PctykKwbT+b8qsrr+iGMSaMxiWQGwCwQm13hpcTfHu/f5S8ei1w733R9hj2o7APvzgg64u8AlW7xa2ZZdWCGGOfaLOH9BePIx3pRT7/D8P120HQXibCAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;squad&quot; title=&quot;squad&quot; src=&quot;/static/7d77e25ad2b5dc1fc2501303113ef72f/2bef9/squad.png&quot; srcSet=&quot;/static/7d77e25ad2b5dc1fc2501303113ef72f/6f3f2/squad.png 256w,/static/7d77e25ad2b5dc1fc2501303113ef72f/01e7c/squad.png 512w,/static/7d77e25ad2b5dc1fc2501303113ef72f/2bef9/squad.png 1024w,/static/7d77e25ad2b5dc1fc2501303113ef72f/71c1d/squad.png 1536w,/static/7d77e25ad2b5dc1fc2501303113ef72f/a878e/squad.png 2048w,/static/7d77e25ad2b5dc1fc2501303113ef72f/f37ba/squad.png 2216w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위 그림은 SQuAD 2.0으로,  답에 해당하는 문구를 찾기 위해 다음 과정을 수행한다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;답이 있는지 없는지 판단한다. 답이 있다면 다음으로 넘어가고, 그렇지 않으면 종료한다.&lt;ul&gt;&lt;li&gt;질문과 문단을 종합적으로 보고 판단하므로, CLS토큰을 활용한다.&lt;ul&gt;&lt;li&gt;질문과 문단을 concat해서 BERT로 인코딩하여 CLS 토큰을 얻는다.&lt;/li&gt;&lt;li&gt;주어진 질문과 문단 쌍에서 답이 없다면(ground truth가 No Answer라면) CLS 토큰을 이진분류하는 output layer에 통과시켜, 크로스 엔트로피로 학습하게 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;CLS 토큰을 통해 답이 있는지 없는지부터 예측 하고, 답이 있는 경우에만 다음 단계로 넘어간다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;단어를 벡터로 변환한다.&lt;/li&gt;&lt;li&gt;각 벡터를 scalar 값으로 변환시켜주는 Fully Connected Layer에 통과시킨다(이 Layer는 모든 벡터에 공통이다)&lt;ul&gt;&lt;li&gt;이 때 Fully Connected Layer의 파라미터가 Random Initialization부터 fine-tuning되는 파라미터이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;단어에 해당하는 스칼라 값들을 모두 softmax에 통과시켜 ground truth의 첫번째에 해당하는 단어를 찾는다.(가장 확률이 높은 단어를 찾으면 된다.)&lt;ul&gt;&lt;li&gt;Softmax-with-loss로 학습시킨다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;정답이 끝나는 지점을 예측할 Fully Connected Layer를 새로 만들고,  Ending word의 포지션을 예측시킨다.(3-4와 동일)&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;SWAG&lt;/code&gt;&lt;/strong&gt; 에서는 주어진 문장이 있을 때, 다음에 나타날 문장을 객관식으로 고르는 형태의 QA다. 이 경우에도 CLS 토큰을 사용한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;4가지 선택지를 모두 하나씩 질문과 concat하여 벡터화하고, Fully Connected Layer를 통과시켜 예측 스칼라값으로 만든다.&lt;/li&gt;&lt;li&gt;4개의 스칼라값을 모두 softmax에 통과시켜, 정답에 해당하는 값의 확률을 높이도록 학습시킨다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;ablation-study&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#ablation-study&quot; aria-label=&quot;ablation study permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Ablation Study&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://fintecuriosity-11.tistory.com/73&quot;&gt;알고리즘이나 모델의 feature를 제거하면서, 그 행위가 성능에 끼치는 영향을 평가하는 방식&lt;/a&gt;을 Ablation Study라고 한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;Big models help a lot&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;Layer를 점점 더 쌓고, 파라미터를 늘릴수록, 즉 Big model일수록 더 좋더라는 이야기.&lt;/li&gt;&lt;li&gt;데이터셋이 3600개밖에 없더라도 파라미터를 110M→340M 개로 늘리니까 더 좋아졌다.&lt;/li&gt;&lt;li&gt;GPU 리소스를 늘리면 늘릴수록 더 높아지더라. 점근선(asymptote)의 형태가 아니더라! 최대한 많이 리소스를 늘려라.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;advanced-self-supervised-pre-training-models&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#advanced-self-supervised-pre-training-models&quot; aria-label=&quot;advanced self supervised pre training models permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Advanced Self-supervised Pre-training Models&lt;/h1&gt;&lt;p&gt;GPT-1과 BERT 이후에 나온 자기지도 사전학습 모델을 알아보자.&lt;/p&gt;&lt;h2 id=&quot;gpt-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gpt-2&quot; aria-label=&quot;gpt 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GPT-2&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;GPT-2&lt;/code&gt;&lt;/strong&gt;는 GPT-1에 비해 다음과 같은 점이 발전되었다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Layer를 훨씬 더 많이 쌓았다. 다만 학습 방식은 Language Modeling으로 동일하다.&lt;/li&gt;&lt;li&gt;학습 데이터가 40GB로 늘어났다. 게다가 데이터의 퀄리티도 좀 더 신경써서 준비했다고 한다.&lt;/li&gt;&lt;li&gt;Language Model이 메인 task 이외의 down-stream task들을 zero-shot setting, 즉 파라미터나 아키텍쳐 조작을 전혀 하지 않고도 수행가능하다는 것을 입증했다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;모티베이션이 된 논문은 다음과 같다.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://arxiv.org/pdf/1806.08730.pdf&quot;&gt;The Natural Language Decathlon : Multitask Learning as Question Answering&lt;/a&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;모든 종류의 자연어 처리 task들이 Question-Answering task로서 처리될 수 있다&lt;/strong&gt;는 점을 시사하였다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;데이터셋&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%8D%B0%EC%9D%B4%ED%84%B0%EC%85%8B&quot; aria-label=&quot;데이터셋 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;데이터셋&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;Reddit&lt;ul&gt;&lt;li&gt;외부링크를 포함한 모든 웹 텍스트를 수집했다.&lt;/li&gt;&lt;li&gt;총 45M개의 링크를 사용했다.&lt;ul&gt;&lt;li&gt;사람이 직접 큐레이팅/필터링한 웹페이지만 수집하였다.&lt;/li&gt;&lt;li&gt;3개 이상의 추천을 받은 게시글들만 수집하였다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;8M 크기의 위키피디아 문서&lt;/li&gt;&lt;li&gt;링크로부터 content를 추출하기 위하여 dragnet과 newspaper를 사용했다고 한다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;전처리&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A0%84%EC%B2%98%EB%A6%AC&quot; aria-label=&quot;전처리 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;전처리&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Byte Pair Encoding(BPE)&lt;/code&gt;&lt;/strong&gt;를 사용하였다.&lt;/p&gt;&lt;h3 id=&quot;gpt-1-대비-구조상의-변화&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gpt-1-%EB%8C%80%EB%B9%84-%EA%B5%AC%EC%A1%B0%EC%83%81%EC%9D%98-%EB%B3%80%ED%99%94&quot; aria-label=&quot;gpt 1 대비 구조상의 변화 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;(GPT-1 대비) 구조상의 변화&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:224px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:179.4642857142857%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;gpt-2&quot; title=&quot;gpt-2&quot; src=&quot;/static/3ac1124d41c7ce6a6de54211ce0ef74c/80b2d/gpt-2.png&quot; srcSet=&quot;/static/3ac1124d41c7ce6a6de54211ce0ef74c/80b2d/gpt-2.png 224w&quot; sizes=&quot;(max-width: 224px) 100vw, 224px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Layer Normalization이 각 sub-block의 입력쪽으로 옮겨졌다(pre-activation residual network와 비슷한 형태).&lt;/li&gt;&lt;li&gt;마지막 self-attention block에서 Layer Normalization이 하나 추가되었다.&lt;/li&gt;&lt;li&gt;각 Layer를 random initialization할 때, Layer의 깊이에 비례해서 값이 더 작아지도록 초기화하였다. 따라서 위쪽에 있는 Layer가 하는 역할이 좀 더 축소되도록 하였다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;zero-shot-setting&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#zero-shot-setting&quot; aria-label=&quot;zero shot setting permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Zero-Shot Setting&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Zero-Shot Setting&lt;/code&gt;&lt;/strong&gt;이란, &lt;strong&gt;데이터가 없이도 해당 task의 수행을 할 수 있음&lt;/strong&gt;을 일컫는 용어이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Question Answering&lt;ul&gt;&lt;li&gt;Task별 학습을 시키지 않고 pre-training만 시킨 GPT-2에 질문(Question)을 준 뒤, 다음에 나올 문장을 예측하라고 했더니, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;F&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;F_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;F&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 스코어가 55가 나왔다고 한다. 잘 학습된 BERT 모델이 89의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;F&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;F_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;F&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;스코어를 기록하는 것을 생각해보면, 학습도 시키지 않은 것 치고는 준수한 성능이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Summarization&lt;ul&gt;&lt;li&gt;CNN과 일간 메일 데이터셋에서 TL;DR(Too Long, Didn&amp;#x27;t Read) 이라는 어구를 기준으로 요약이 있었는데, 이 때문에 TL;DR을 GPT-3에 입력으로 주면 요약을 해 준다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Translation&lt;ul&gt;&lt;li&gt;Summarization과 같은 방식으로, [In French, In Korean, ...]이라는 어구가 있으면 번역까지 수행해준다고 한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;gpt-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gpt-3&quot; aria-label=&quot;gpt 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GPT-3&lt;/h2&gt;&lt;p&gt;가장 최근에 나온 GPT 모델로, 모델의 구조에 변화가 있었다기보다는, &lt;strong&gt;이전과 비교할 수 없을 정도의 attention block을 쌓아 파라미터수를 어마어마하게 많이 늘렸다(150B)&lt;/strong&gt;. 또, &lt;strong&gt;배치사이즈도 3.2M정도가 되도록 최대한 키우자 더 좋은 성능&lt;/strong&gt;을 보였다고 한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:167px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:180.8383233532934%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAkCAIAAAAGkY33AAAACXBIWXMAAAsSAAALEgHS3X78AAAD4ElEQVRIx6WVh5LjNgyG9f6Pl2xu3eQmW91qLJKo4nykfLazezOXTDAYCoQIEgT/n/SCILher+fzue/7+38U73K5RFFEsJSye5O2bZluaenSaq3pLr+mabLBYRQFwWW/3/u+v16v94fD6XTCPh6Pu92OLgZdEtw5Wa1Wm83GGGODkziKWToMb7dbEsfYtHmWheH1erlkWRacz5h4+Of+h1VVPdLWw9wOk5nmYZ7H+32Y766dnccatGa+9+M0zvPXPR8Ldam1n4tVUn8m9TZvPuMKY38TKMYuE6lso0bLrqcuZVmyclEUZO4FpQ6b9lhKhp5KhX2p1LXWsWgTQYzVTHXYqh9arYUQ1IxIauadS3Vt9CZrVml9KmUi21x1jM6kjVk0VV0sW9kbrVTTNEupbdrtMHbj6DbJzqceHX+lRHzfs23wzr+X6Zt4SikplZBS/wsBHku7gMfj3A+HA+38Lavfw/NWFGcHb6qf3255nqdpmjnB4GDwA6LFQ7WwKfgL26v12t8DxMPeif8Tqi+Qur/b7ZYu2HxSyBOiqeuK09du91IIDMESdW0Ppuan/YtHOGHUW9rtWHboULRD0Q1VN2LQrXrn7MaqG4SZGjMCWA5lHEfjBMPCM6jVoZA/kupHXPl582dk4Xks5ALPTVaDnIiEHMIWeMIiyu4QVmuAyehdLkAbUN/lDd1DAbwlqM51R7w2I8wmkrBH2udSE0Aweq7Atg5rzYxMEVQKA5ADWBc8GGO58SrYkvY6bf4IS5ZinVRaMKcW220mrUEkk5I2JeTAXsGFNlVrZD+gTW9Eb5run9o77XozPvjwhJN3/x/i9bbsbMZQBmVFW8W2X2co/XQ+hXGw2ls7+fz8fLL0XZ4ZLsRajNGJZRXnBrC52RaAIaDLIe2BJ9YBZnhcXnbBV9phGAJa0AtuAbDvBM/GCTfudrej6/t7uoxh5DAMP6ud50kScwBF4TiVZ6AZQsXuAsbD7BALT5Ikmb2A4+finjSzHuduvreTVWdM7WQ9nfXMnCktv8ZvfHcgqRRX70dUfsTlNrNoAeSHmwDnf0UVV7IDiZKdoSzwnzKRAvv3gsrCE2IAZqa4OlQCci5jnP5NQAmHba3MwDNF8SjkUvklmHvbEovgS60uLhGCF2IxHdiOhRb2teuWI3gEW2LUep3Wjn0NVzfEwECX6baZ4MYH4aL9+gY/glnBty+GZSUtZIKM29xS0nct3BJt944Wl7YL9l0wg5Y8SR77ZDNXblK1UPIdc6+HjiQ/ooqdx+6VIcl0MdTj0bHV7gf1hZKWjGbQw6jMSPsLdX5KPfDguIfiufjfFXQAjweNaoMAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;gpt-3-few-shot&quot; title=&quot;gpt-3-few-shot&quot; src=&quot;/static/d51200945978aec912e03f6011da3c6e/21521/gpt-3-few-shot.png&quot; srcSet=&quot;/static/d51200945978aec912e03f6011da3c6e/21521/gpt-3-few-shot.png 167w&quot; sizes=&quot;(max-width: 167px) 100vw, 167px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;GPT-2에 비해 눈에 띄는 특징은 다음과 같다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;GPT-2에서는 &amp;#x27;가능성&amp;#x27;정도로만 보였던 &lt;code&gt;Zero-shot setting&lt;/code&gt;이 놀라운 수준으로 발전하였다.&lt;ul&gt;&lt;li&gt;학습에서 전혀 활용하지 않았던 텍스트를 translation 했을때도 정상적으로 기능한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;하고자 하는 task(예를 들자면 번역)를 주고, 예시를 주면, 자연어 생성 task로 인식하여 정확도를 평가하고 스스로 학습한다. 이를 &lt;strong&gt;&lt;code&gt;One-shot&lt;/code&gt;&lt;/strong&gt;이라고 한다. 데이터를 단 한 쌍(예시)만 주었다는 말이다.&lt;ul&gt;&lt;li&gt;신기한 점은, 모델 자체의 파라미터를 변경시켜가며 학습한 것이 아니라, 데이터를 input 텍스트의 일부로서 제시했는데도 task를 수행했다는 것이다!&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;동일한 맥락으로, 몇 개 안되는 예시 데이터를 주고 task를 수행하도록 하는 Few-shot이 가능해졌다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:60.9375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;gpt-3-few-shot-performance&quot; title=&quot;gpt-3-few-shot-performance&quot; src=&quot;/static/60116e0b144deab5846ef610322313da/2bef9/gpt-3-few-shot-performance.png&quot; srcSet=&quot;/static/60116e0b144deab5846ef610322313da/6f3f2/gpt-3-few-shot-performance.png 256w,/static/60116e0b144deab5846ef610322313da/01e7c/gpt-3-few-shot-performance.png 512w,/static/60116e0b144deab5846ef610322313da/2bef9/gpt-3-few-shot-performance.png 1024w,/static/60116e0b144deab5846ef610322313da/e996b/gpt-3-few-shot-performance.png 1050w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;연구 결과에 따르면, 모델 사이즈를 키우면 키울수록, &lt;strong&gt;&lt;code&gt;Zero/One/Few shot&lt;/code&gt;&lt;/strong&gt;의 성능이 계속해서 오른다고 한다.&lt;/p&gt;&lt;h2 id=&quot;albert&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#albert&quot; aria-label=&quot;albert permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;ALBERT&lt;/h2&gt;&lt;p&gt; &lt;strong&gt;&lt;code&gt;A Lite BERT(ALBERT)&lt;/code&gt;&lt;/strong&gt;는  기존의 BERT를 경량화 시킨 모델이다. GPT와 같이 모델이 굉장히 거대해지고 리소스와 연산량이 많아지는 형태와는 달리, 오히려 복잡했던 BERT를 개선하는 데에 집중했다. 모델 사이즈를 줄이고, 학습시간과 리소스를 줄이면서도 성능은 크게 떨어뜨리지 않는 경량화 형태의 Pre-trained Model이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;BERT에서 큰 버젼과 작은 버젼이 있듯이, ALBERT에도 모델 파라미터 사이즈에 따라 좀 더 큰 모델과 작은 모델이 있다. 물론 더 큰 모델을 사용할 때 좀 더 좋은 성능을 낸다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이해하기 쉽게 잘 정리된 글이 있으니 아래의 내용들이 이해가 가지 않는다면 이 링크를 참조하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://y-rok.github.io/nlp/2019/10/23/albert.html&quot;&gt;[ALBERT 논문 Review] ALBERT: A LITE BERT FOR SELF-SUPERVISED LEARNING OF LANGUAGE REPRESENTATIONS&lt;/a&gt;&lt;/p&gt;&lt;h3 id=&quot;factorized-embedding-parameterization&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#factorized-embedding-parameterization&quot; aria-label=&quot;factorized embedding parameterization permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Factorized Embedding Parameterization&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:41.79687500000001%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;factorized-embedding-parameterization&quot; title=&quot;factorized-embedding-parameterization&quot; src=&quot;/static/d1f32130e74a44018411d7a41f19984d/2bef9/factorized-embedding-parameterization.png&quot; srcSet=&quot;/static/d1f32130e74a44018411d7a41f19984d/6f3f2/factorized-embedding-parameterization.png 256w,/static/d1f32130e74a44018411d7a41f19984d/01e7c/factorized-embedding-parameterization.png 512w,/static/d1f32130e74a44018411d7a41f19984d/2bef9/factorized-embedding-parameterization.png 1024w,/static/d1f32130e74a44018411d7a41f19984d/71c1d/factorized-embedding-parameterization.png 1536w,/static/d1f32130e74a44018411d7a41f19984d/a878e/factorized-embedding-parameterization.png 2048w,/static/d1f32130e74a44018411d7a41f19984d/e8d55/factorized-embedding-parameterization.png 2894w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;기존의 BERT에서 Embedding vector 사이즈 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;E&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 hidden vector size &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;H&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 항상 같아야했다. 여러 Attention Block을 쌓기 때문에, 같은 크기로 들어가고 나가야 다음 블록에 동일한 형태로 전달될 수 있다.&lt;/p&gt;&lt;p&gt;문제는, 단어간의 관계를 인코딩하여 저장해야하므로 많은 정보가 들어가는 dependent 벡터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;H&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 크기를 맞추기 위해, &lt;strong&gt;단어간의 관계를 생각하지 않아도 되는 independent 벡터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;E&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 필요 이상으로 커진다&lt;/strong&gt;는 것이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:42.1875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAICAIAAAB2/0i6AAAACXBIWXMAABYlAAAWJQFJUiTwAAABF0lEQVQY03WRy27DIBBF+f9va1fpoqrUSnZoZHAwmMcAw6sTJ8qquRoWoOFe5sDGGLVWAFCHjDGllNYanfdD47UYrVKqc14sQkq5LIsQQm1bqfXVnacjg4yru0orL9uFX3lMERFzzhSutfbe37rvd8YIAGSeEfvNYjALaVIzN+eznmc1117pJdS6x/p5CV9LwPrI+ZZpWuFXw/vk3yYnLLIQUe678u5qrTB7y3m0Vnuft3Ti8MFDLrf5DZQTDz9rAuxcBa4AcmPOBe8Jkq4tIW73EEqLpdsQHWTatt59atpn49MxQBvH2CxCCtG6tOYSLAhqHIjjP8i9EwfyGc9ihMdal3N5tByfdEdKwElPtuQZwqMAiEz7A7V40Izdz0miAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;albert-fep2&quot; title=&quot;albert-fep2&quot; src=&quot;/static/dc55109428294d55d3be60ea96110ffa/2bef9/albert-fep2.png&quot; srcSet=&quot;/static/dc55109428294d55d3be60ea96110ffa/6f3f2/albert-fep2.png 256w,/static/dc55109428294d55d3be60ea96110ffa/01e7c/albert-fep2.png 512w,/static/dc55109428294d55d3be60ea96110ffa/2bef9/albert-fep2.png 1024w,/static/dc55109428294d55d3be60ea96110ffa/71c1d/albert-fep2.png 1536w,/static/dc55109428294d55d3be60ea96110ffa/a878e/albert-fep2.png 2048w,/static/dc55109428294d55d3be60ea96110ffa/39148/albert-fep2.png 2876w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위 이미지를 보면, BERT에서는 원래 4x1 사이즈의 임베딩 벡터가 H에 맞춰주기 위하여 4x4로 늘어나는 것을 볼 수 있다.&lt;/p&gt;&lt;p&gt;이를 해결하기위해 ALBERT는 &lt;strong&gt;&lt;div&gt;Embedding Matrix를 위의 이미지처럼 두 Matrix의 곱으로 쪼갠다&lt;/div&gt;&lt;/strong&gt;. 가령 H(=4)에 맞추기 위하여 4x4였던 행렬을, 4x2 크기의 행렬로 두고 추가적으로 Layer를 하나 더 두어, Word별로 구성되는 2차원 벡터를 4차원으로 선형변환(W)시켜주도록 한다.  이 때 W는 H의 크기로 변환될 수 있도록 적당한 크기를 가지면 된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;row-rank matrix factorization&lt;/code&gt;&lt;/strong&gt;이라고 하는 기법이다.&lt;/li&gt;&lt;li&gt;위의 예제에서는 잘 와닿지 않지만, H의 크기가 100이고 쪼갠 Matrix의 column length는 10이라고 생각해보자. 파라미터 수가 확 줄어듦이 체감될 것이다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;cross-layer-parameter-sharing&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#cross-layer-parameter-sharing&quot; aria-label=&quot;cross layer parameter sharing permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Cross-layer Parameter Sharing&lt;/h3&gt;&lt;p&gt;Self-attention block들이 가지는 학습 파라미터들에는 무엇들이 있을까?&lt;/p&gt;&lt;ul&gt;&lt;li&gt;임베딩된 input 벡터가 Query, Key, Value 각각의 역할을 하도록 변형시켜주는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;/msup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;/msup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;[W^Q,W^K,W^V]&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0913309999999998em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;Q&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;multi head라면 행렬 세트가 총 head개가 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;Z&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Z^t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7935559999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;Z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7935559999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;들을 concat한 후  다시 원래의 Hidden State Vector 크기의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;Z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 줄여주기 위한 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^O&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8413309999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;물론, 각각의 Self Attention block마다 이 파라미터값은 모두 다를것이다.&lt;/p&gt;&lt;p&gt;그런데 ALBERT는 서로 다른 Layer, 즉 &lt;strong&gt;서로 다른 Self-Attention Block에 존재하는 파라미터들을 서로 공유&lt;/strong&gt;하는 방법을 제시한다. 이를 &lt;strong&gt;&lt;code&gt;Cross-layer Parameter Sharing&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:33.984375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA5UlEQVQY012R2aqDQBBE8/8f5oM++WRycd/XCK4ZRbyHaZNAChyK6q7qnvF2nqd6qWVZjuPYtm0Yhmma5nnu+14pNWhQFV0IJyWMN76qqhzHeWq4ruv7fpZlhmE0TXMHzp2GMAzzPC/LMgiCKIqKorjMbdsi1XUNoSOOY1Isy+JEJ4st8LcaEKJJ+TUzioJw0zTRPc97/D3QSUzTlIEQtkiS5Gsm/vkG2YyybZubSxa3lbFUOcmi5zIjdV0nKjZWgmOjA102krXpzDXGcbzML411XeUZGbho8LCrxkehTf7Fvu8Y/wHKzYSS/L84ogAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;albert-sharing&quot; title=&quot;albert-sharing&quot; src=&quot;/static/a3bac524d238aad508f7ecf09fa6c8ab/2bef9/albert-sharing.png&quot; srcSet=&quot;/static/a3bac524d238aad508f7ecf09fa6c8ab/6f3f2/albert-sharing.png 256w,/static/a3bac524d238aad508f7ecf09fa6c8ab/01e7c/albert-sharing.png 512w,/static/a3bac524d238aad508f7ecf09fa6c8ab/2bef9/albert-sharing.png 1024w,/static/a3bac524d238aad508f7ecf09fa6c8ab/71c1d/albert-sharing.png 1536w,/static/a3bac524d238aad508f7ecf09fa6c8ab/263f4/albert-sharing.png 1996w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Shared-FFN&lt;/code&gt;&lt;/strong&gt; : Layer 간에 feed-forward network의 파라미터만 공유한다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Shared-attention&lt;/code&gt;&lt;/strong&gt; : Layer 간에 attention 파라미터들만 공유한다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;All-shared&lt;/code&gt;&lt;/strong&gt; : 둘 다 공유한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이처럼 파라미터를 공유하였을 때, &lt;strong&gt;파라미터의 개수는 크게 줄었음에도 불구하고 모델의 성능은 그다지 크게 떨어지지 않았음을 입증&lt;/strong&gt;하였다.&lt;/p&gt;&lt;h3 id=&quot;sentence-order-prediction&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#sentence-order-prediction&quot; aria-label=&quot;sentence order prediction permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Sentence Order Prediction&lt;/h3&gt;&lt;p&gt;BERT 이후의 후속연구에서, BERT 모델이 기존에 pretraining하던 &lt;code&gt;Next Sentence Prediction task&lt;/code&gt;는 사실 너무 쉬워서 그다지 실효성이 없는 것으로 드러났다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;두 문장의 출처가 다르다면, 사실상 전혀 다른 내용일 확률이 높다.&lt;/li&gt;&lt;li&gt;따라서 그냥 비슷한 단어나 문맥이 많이 등장했는가 정도로 선후관계를 파악하게 된다.&lt;/li&gt;&lt;li&gt;이는 선후 관계보다는 topic prediction에 가깝다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;ALBERT에서는 해당 task의 pretraining을 빼고 좀 더 유의미한 task들을 집어넣어, 모델의 성능을 확장했다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;두 독립적인 문장을 가져와 선후관계를 파악하는 것이 아니라, 항상 연속적인 두 문장을 가져온다.&lt;/li&gt;&lt;li&gt;그 문장을 원래의 순서대로 concat했을 때 정방향으로 예측하고, 역순으로 concat했을 때 역방향으로 예측하도록 학습시킨다.(이진분류)&lt;/li&gt;&lt;li&gt;이를 &lt;code&gt;negative sample&lt;/code&gt;이라고 하는데, 인접 문장이므로 순서와 관계없이 비슷한 단어가 당연히 많이 등장한다.&lt;ul&gt;&lt;li&gt;따라서 정말로 논리적인 흐름을 주의깊게 파악해야 task를 풀 수 있는 pretraining 형태가 되었다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:14.453125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAADCAIAAAAcOLh5AAAACXBIWXMAABYlAAAWJQFJUiTwAAAAdklEQVQI1zWOWQ7DIAxEc/+DcQl2BEiACoGwfLYjSN+HNczYxtec895orTnn1trWWs55jIHIew+fMWaMef6stb6bSymFzDknhJBS4olKCKGUwoQIIUBjy2kA6H+Ha61nWe8dAifAwW8ppXMOohgjKtJSymdzhn+Yk6TUOcINogAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;albert-sop&quot; title=&quot;albert-sop&quot; src=&quot;/static/bbb2c2706f67613d7b2146849ebe59b7/2bef9/albert-sop.png&quot; srcSet=&quot;/static/bbb2c2706f67613d7b2146849ebe59b7/6f3f2/albert-sop.png 256w,/static/bbb2c2706f67613d7b2146849ebe59b7/01e7c/albert-sop.png 512w,/static/bbb2c2706f67613d7b2146849ebe59b7/2bef9/albert-sop.png 1024w,/static/bbb2c2706f67613d7b2146849ebe59b7/71c1d/albert-sop.png 1536w,/static/bbb2c2706f67613d7b2146849ebe59b7/772aa/albert-sop.png 2042w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;논문에 첨부된 위의 실험결과를 보면, Next Sentence Prediction(NSP)를 사용했을때는 아예 사용하지 않았을 때와 별 차이가 없거나 오히려 성능이 떨어지기까지 한다. 이에 비해 &lt;strong&gt;&lt;code&gt;Sentence Order Prediction(SOP)&lt;/code&gt;&lt;/strong&gt;는 좀 더 좋은 개선된 성능을 보여주고 있다.&lt;/p&gt;&lt;h2 id=&quot;electra&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#electra&quot; aria-label=&quot;electra permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;ELECTRA&lt;/h2&gt;&lt;p&gt;2020년에 발표된 논문으로, GPT의 standard한 LM이나, BERT의 Masked LM task에서 나아가, &lt;strong&gt;GAN 형태의 모델링&lt;/strong&gt;을 제시하고 있다.&lt;/p&gt;&lt;p&gt;MLM(Masked language Modeling)을 통해 마스킹된 단어를 복원해주는 모델-&lt;strong&gt;&lt;code&gt;Generator&lt;/code&gt;&lt;/strong&gt;를 하나 두고, 또 Generator가 복원한 단어들을 받아 이 단어가 원본인지 또는 generator에 의해 복원된 단어인지를 예측하는 모델-&lt;strong&gt;&lt;code&gt;Discriminator&lt;/code&gt;&lt;/strong&gt;를 둔다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;generator는 BERT의 형태로 볼 수 있다.&lt;/li&gt;&lt;li&gt;이 때 Generator와 Discriminator이라는 구조를 Generative Adversarial Network(GAN)형태로 볼 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이렇게 모델 학습을 진행할 경우에, Pre-train된 모델로서 사용할 수 있는 부분이 Generator와 Discriminator 두 부분이 된다. 이 중 &lt;strong&gt;&lt;code&gt;ELECTRA&lt;/code&gt;&lt;/strong&gt;는 Discriminator를 가져다가 downstream task에 맞게 fine-tuning하여 사용하는 방식이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:40.625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAICAIAAAB2/0i6AAAACXBIWXMAABYlAAAWJQFJUiTwAAAA6klEQVQY03VQi26EIBC8///H9lrTqOAJCvvAB3qj9GWaTgJMdneW3bntJ5g5EM3TrCx994gheudCiK53RMw4xJomUY0gqkV1Kw+iSXXo3evL3dStsMYxvt2rzvYk6ruOArWma42tG4OmP+Kcs1B8GPtRN96Pa84ll7eDqIh3HmQ7ATLPcyGHWFUwnu16RPezCmf/gogOw1iqCy7idV1DJOwFsv9BSglz/isGnB+BzxwmYZYT8AZxf/15mqaLmFhiJETxj2lM9V59w1qLFDbXBE8VNWh6cXv7hXQWld64ixEgx2LMcHdZlqJ6ArLS0RgfeAZNAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;electra-performance&quot; title=&quot;electra-performance&quot; src=&quot;/static/30a0c7ee8b543008c320d70db20172b3/2bef9/electra-performance.png&quot; srcSet=&quot;/static/30a0c7ee8b543008c320d70db20172b3/6f3f2/electra-performance.png 256w,/static/30a0c7ee8b543008c320d70db20172b3/01e7c/electra-performance.png 512w,/static/30a0c7ee8b543008c320d70db20172b3/2bef9/electra-performance.png 1024w,/static/30a0c7ee8b543008c320d70db20172b3/71c1d/electra-performance.png 1536w,/static/30a0c7ee8b543008c320d70db20172b3/a878e/electra-performance.png 2048w,/static/30a0c7ee8b543008c320d70db20172b3/8ad1b/electra-performance.png 2320w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;ELECTRA의 논문에 따르면, &lt;strong&gt;대부분의 BERT 모델보다 동일한 학습량 대비 성능이 더 좋다&lt;/strong&gt;고 한다.&lt;/p&gt;&lt;h2 id=&quot;light-weight-models&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#light-weight-models&quot; aria-label=&quot;light weight models permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Light-weight Models&lt;/h2&gt;&lt;p&gt;이외에도 최근의 모델 연구동향은 경량화 기술에 주목하고 있다. GPT-3같은 대형 모델의 성능은 놀랍지만, 연구나 활용을 하기에는 오히려 접근이 힘들다는 문제가 있다.&lt;/p&gt;&lt;p&gt;기존의 정확도를 유지하면서도 파라미터수나 레이어 수를 줄임으로써 모델의 크기와 학습속도를 빠르게 하려는 노력들이 이어지는 중이다. 이를 통해 클라우드나 서버를 통한 AI 적용이 아니라 휴대폰이나 IOT기기에서도 딥러닝이 가능하게 할 수 있다.&lt;/p&gt;&lt;h3 id=&quot;distillbert&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#distillbert&quot; aria-label=&quot;distillbert permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;DistillBERT&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://huggingface.co/&quot;&gt;Huggingface&lt;/a&gt;에서 2019년도 발표한 모델로,  Teacher 모델과 Student 모델로 이루어져있다. 큰 사이즈의 Teacher 모델이 기존의 방식으로 학습을 수행한 후, Student 모델은 더 적은 파라미터로 Teacher 모델의 수행방식을 모사하여 비슷한 성능을 낸다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Teacher 모델이 예측한 결과를 Student 모델이 softmax에 주는 ground-truth로 삼아 학습한다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;tinybert&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#tinybert&quot; aria-label=&quot;tinybert permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;TinyBERT&lt;/h3&gt;&lt;p&gt;2020년 발표된 &lt;code&gt;TinyBERT&lt;/code&gt; 역시 BERT를 경량화시킨 모델이지만, 좀 더 발전된 형태이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;DistillBERT과 마찬가지로 target distribution을 모사한 뒤, Student 모델에서 이를 ground-truth로 삼고 학습한다.&lt;/li&gt;&lt;li&gt;추가적으로, Student 모델이 Teacher 모델의 Attention Matrix와 Hidden state Vector까지도 유사하도록 모사한다.&lt;ul&gt;&lt;li&gt;최종 결과물뿐만 아니라 &lt;strong&gt;중간 결과물까지도 비슷하도록 모사&lt;/strong&gt;한다고 볼 수 있다.&lt;/li&gt;&lt;li&gt;MSE-Loss를 사용한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 때, Student 모델의 벡터 크기가 일반적으로 Teacher 모델의 벡터보다 작기 때문에 이를 욱여넣는 과정에서 전정보의 손실이 발생할 수 있다.&lt;/p&gt;&lt;p&gt;이를 방지하고자 TinyBERT는 벡터간의 크기 변환 과정에 Fully Connected Layer를 두어, 기존의 정보를 어느정도 유지하면서 더 적은 벡터로 이를 모사할 수 있게 된다.&lt;/p&gt;&lt;h2 id=&quot;fusing-knowledge-graph-into-language-model&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#fusing-knowledge-graph-into-language-model&quot; aria-label=&quot;fusing knowledge graph into language model permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Fusing Knowledge Graph into Language Model&lt;/h2&gt;&lt;p&gt;기존의 Pritrained 모델과 지식(knowledge) 그래프를 조합한 최신 연구동향이다.&lt;/p&gt;&lt;p&gt;BERT가 2018년도에 등장한 이후, BERT가 정말로 언어적인 이해를 제대로 하고 있는 것인지 분석하는 연구들이 이어지면서, BERT가 데이터셋에 나와있지 않은 종류의 문장들은 잘 이해하지 못한다는것이 드러났다.&lt;/p&gt;&lt;p&gt;주어진 문장에서 나타나는 지식 뿐만 아니라, 외부 지식(또는 상식)도 자연어 처리에서 중요한데, 이를 연구하는 분야가 &lt;strong&gt;&lt;code&gt;Knowledge Graph&lt;/code&gt;&lt;/strong&gt; 분야이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;개체간의 관계, 개념등을 잘 정의하고, 정형화해서 만들어 둔 것을 Knowledge Graph라고 한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Knowledge graph를 어떻게 BERT등에 적용시킬 수 있을까에 대한 연구 동향으로 ERNIE, KagNET등이 있다.&lt;/p&gt;&lt;hr/&gt;&lt;h3 id=&quot;reference&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#reference&quot; aria-label=&quot;reference permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Reference&lt;/h3&gt;&lt;p&gt;논문&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/pdf/2005.14165.pdf&quot;&gt;Language Models are Few-Shot Learners&lt;/a&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/pdf/1909.11942.pdf&quot;&gt;ALBERT: A LITE BERT FOR SELF-SUPERVISED LEARNING OF LANGUAGE REPRESENTATIONS&lt;/a&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://openreview.net/pdf?id=r1xMH1BtvB&quot;&gt;ELECTRA: PRE-TRAINING TEXT ENCODERS AS DISCRIMINATORS RATHER THAN GENERATORS&lt;/a&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/pdf/1910.01108.pdf&quot;&gt;DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter&lt;/a&gt;&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://arxiv.org/pdf/1909.10351.pdf&quot;&gt;TinyBERT: Distilling BERT for Natural Language Understanding&lt;/a&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;a href=&quot;https://openai.com/blog/language-unsupervised/&quot;&gt;Improving Language Understanding with Unsupervised Learning&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://rajpurkar.github.io/SQuAD-explorer/&quot;&gt;The Stanford Question Answering Dataset&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://blog.floydhub.com/gpt2/&quot;&gt;GPT-2: How to Build &amp;quot;The AI That&amp;#x27;s Too Dangerous to Release&amp;quot;&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://openai.com/blog/better-language-models/&quot;&gt;Better Language Models and Their Implications&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;http://jalammar.github.io/illustrated-transformer/&quot;&gt;The Illustrated Transformer&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[NLP 07 - 'Attention is all you need'에 나온 Transformer 모델 알아보기]]></title><description><![CDATA[Transformer by 주재걸 교수님, BoostCamp AI Tech 4주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/28_transformer/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/28_transformer/</guid><pubDate>Thu, 18 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;transformer&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#transformer&quot; aria-label=&quot;transformer permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Transformer&lt;/h1&gt;&lt;p&gt;기존에는 Attention 모듈이 RNN이나 CNN 모듈의 Add-on 모듈로 사용되어왔다.&lt;/p&gt;&lt;p&gt;그러나 2017년 발표된 논문 &lt;a href=&quot;https://arxiv.org/pdf/1706.03762.pdf&quot;&gt;&amp;#x27;Attention is all you need&amp;#x27;&lt;/a&gt;는 기존의 RNN과 CNN을 모두 걷어내고, &lt;strong&gt;오로지 Attention만으로 구축하여 시퀀스 데이터를 입출력할 수 있는 Transformer 모델을 구축&lt;/strong&gt;하였다.&lt;/p&gt;&lt;h2 id=&quot;기존-rnn-모델의-한계&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B8%B0%EC%A1%B4-rnn-%EB%AA%A8%EB%8D%B8%EC%9D%98-%ED%95%9C%EA%B3%84&quot; aria-label=&quot;기존 rnn 모델의 한계 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;기존 RNN 모델의 한계&lt;/h2&gt;&lt;p&gt;RNN은 정보가 여러 time step을 거치면서, 멀리있는 정보가 유실/변질 되는 long-term dependecy 문제가 있었다. &lt;/p&gt;&lt;h3 id=&quot;bi-directional-rnns&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#bi-directional-rnns&quot; aria-label=&quot;bi directional rnns permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Bi-Directional RNNs&lt;/h3&gt;&lt;p&gt;기존 RNN은 단방향으로 정보를 전달하기 때문에 어느정도 먼 거리의 정보를 반영하기 어려웠는데, 이를 해결하고자 &lt;strong&gt;양방향으로 RNN을 병렬 구성한 &lt;code&gt;Bi-Directional RNN&lt;/code&gt;&lt;/strong&gt;이 나오게 되었다.(Forward RNN, Backward RNN)&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:47.265625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;bidirectional-rnn&quot; title=&quot;bidirectional-rnn&quot; src=&quot;/static/21276ea64553966f756d245fb319eec5/2bef9/bidirectional-rnn.png&quot; srcSet=&quot;/static/21276ea64553966f756d245fb319eec5/6f3f2/bidirectional-rnn.png 256w,/static/21276ea64553966f756d245fb319eec5/01e7c/bidirectional-rnn.png 512w,/static/21276ea64553966f756d245fb319eec5/2bef9/bidirectional-rnn.png 1024w,/static/21276ea64553966f756d245fb319eec5/71c1d/bidirectional-rnn.png 1536w,/static/21276ea64553966f756d245fb319eec5/a878e/bidirectional-rnn.png 2048w,/static/21276ea64553966f756d245fb319eec5/20f38/bidirectional-rnn.png 2818w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;이 때 동일한 time step에서의 hidden state vector &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h^f_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.2127719999999997em;vertical-align:-0.24575599999999995em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9670159999999999em&quot;&gt;&lt;span style=&quot;top:-2.454244em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.1809080000000005em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.24575599999999995em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h^b_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.096108em;vertical-align:-0.247em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.849108em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 concat되어 인코딩 벡터를 구성한다.&lt;/p&gt;&lt;h2 id=&quot;transformer-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#transformer-1&quot; aria-label=&quot;transformer 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Transformer&lt;/h2&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:74.609375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAPCAIAAABr+ngCAAAACXBIWXMAABYlAAAWJQFJUiTwAAABa0lEQVQoz5WTTW+CQBCG+//P7cmkaQ8m7aGpbYyI31VkFRTFLwRFEFkEpQSUCkK1RTnWir6nycw8++5Odm5+zsrZOKoy9z3vZPXmPCxAEW8BuFSCeL/fXwf3BlwKLbJj4VLn/UG7IPC2XqmSSSTjZaywWOh/zU87h00b207V2/EyUaD6kjCNhj3PV+fKiKX7XXLC84DAEsmnbB6F8vwcHBZW1qqIg/vkazydxGpEu/hMo3dMDeGFWTRsmWaH4XPNYZOfs6NJDcdyWQQAIEly9MAMw2pTRAF5qH28BNemqokqcktW3qaSEu1sWuseVW+kY1T2kRsyXeydRmNsIy+IcvTAjE+j2xnggByNxeVSF8QZQbaoVgdC+C8cZl3X1XVdU9Uxx32t10Ha33oizysQBnnLtM69OTzC0BZ0scTgIOCnvX4TzUjM0P8+6IIfttstVNU0jONibBRZ9lz3isXwjwoC27Y1TXMc52TbLzznUuM+sCWpAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;transformer_architecture&quot; title=&quot;transformer_architecture&quot; src=&quot;/static/efbf89902cd0d152408930307c8e8459/2bef9/transformer_architecture.png&quot; srcSet=&quot;/static/efbf89902cd0d152408930307c8e8459/6f3f2/transformer_architecture.png 256w,/static/efbf89902cd0d152408930307c8e8459/01e7c/transformer_architecture.png 512w,/static/efbf89902cd0d152408930307c8e8459/2bef9/transformer_architecture.png 1024w,/static/efbf89902cd0d152408930307c8e8459/71c1d/transformer_architecture.png 1536w,/static/efbf89902cd0d152408930307c8e8459/a878e/transformer_architecture.png 2048w,/static/efbf89902cd0d152408930307c8e8459/a6d66/transformer_architecture.png 2070w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h2 id=&quot;인코더-구조&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9D%B8%EC%BD%94%EB%8D%94-%EA%B5%AC%EC%A1%B0&quot; aria-label=&quot;인코더 구조 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;인코더 구조&lt;/h2&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:110.546875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;transformer-diagram&quot; title=&quot;transformer-diagram&quot; src=&quot;/static/a7eba9e684023e26b7be5ddf598c5fb7/2bef9/transformer-diagram.png&quot; srcSet=&quot;/static/a7eba9e684023e26b7be5ddf598c5fb7/6f3f2/transformer-diagram.png 256w,/static/a7eba9e684023e26b7be5ddf598c5fb7/01e7c/transformer-diagram.png 512w,/static/a7eba9e684023e26b7be5ddf598c5fb7/2bef9/transformer-diagram.png 1024w,/static/a7eba9e684023e26b7be5ddf598c5fb7/6937a/transformer-diagram.png 1094w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ol&gt;&lt;li&gt;Input으로 주어진 3개의 벡터(I, go, home)는 각각 본인의 차례에서 선형변환(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^Q&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8413309999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;Q&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)되어 query 벡터로 기능하고, 자기 자신을 포함한 input 벡터들을 선형변환(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^K&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8413309999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)한 key 벡터들과 내적하여 새로운 벡터를 만들어낸다.&lt;ul&gt;&lt;li&gt;이미지에서 [3.8,-0.2,5.9]&lt;/li&gt;&lt;li&gt;이는 각 벡터간의 유사도를 검사하여, input된 단어들간의 관계를 정의한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;이후, 이 벡터는 softmax를 거쳐 확률로 변환하여 가중치 벡터를 만든다.&lt;/li&gt;&lt;li&gt;다시 기존의 input 벡터들을 선형변환(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^V&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8413309999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)한 value 벡터와 가중평균내어 총합이 1이 되는 벡터(attention output vector)를 구성하고, 이것이 곧 해당 input 벡터의 인코딩 벡터가 된다.&lt;ul&gt;&lt;li&gt;구해진 유사도를 바탕으로 가중평균을 내는 과정이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;이 과정을 통해 나온 인코딩 벡터 값으로 여러 input 중 어떤 input에 어느 정도 비율로 집중(attention)해야 할 지 알 수 있게 된다.&lt;/p&gt;&lt;div&gt;&lt;p&gt;이 때, 선형변환(W)을 거치지 않고 input 벡터를 그대로 사용하면, 당연하게도 자기 자신과의 내적이 가장 커지기 때문에, 결과로 나온 h에서는 자기자신에 대한 attention이 가장 커져 정보의 유효성이 부족해지는 문제가 생길 수 있다.&lt;/p&gt;&lt;/div&gt;&lt;p&gt;이러한 transformer의 구조는 &lt;strong&gt;정보가 위치한 거리와 관계없이 유사도를 측정하여 attention을 분배함&lt;/strong&gt;으로써 기존의 RNN구조가 가지는 &lt;strong&gt;&lt;div&gt;Long-term Dependency 문제를 근본적으로 해결&lt;/div&gt;&lt;/strong&gt;했다고 할 수 있다.&lt;/p&gt;&lt;h3 id=&quot;뜯어보기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%9C%AF%EC%96%B4%EB%B3%B4%EA%B8%B0&quot; aria-label=&quot;뜯어보기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;뜯어보기&lt;/h3&gt;&lt;p&gt;Transformer 모델에서 핵심적 역할을 하는 세 벡터, &lt;strong&gt;&lt;code&gt;Query&lt;/code&gt;&lt;/strong&gt;(이하 &lt;strong&gt;Q&lt;/strong&gt;)와 &lt;strong&gt;&lt;code&gt;Key&lt;/code&gt;&lt;/strong&gt;(이하 &lt;strong&gt;K&lt;/strong&gt;), &lt;strong&gt;&lt;code&gt;Value&lt;/code&gt;&lt;/strong&gt;(이하 &lt;strong&gt;V&lt;/strong&gt;)에 대해서 살펴보자.&lt;/p&gt;&lt;p&gt;Output은 V들의 가중평균인데, 이 가중평균은 결국 &lt;strong&gt;Q와 K의 내적&lt;/strong&gt;으로 구성된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때,  &lt;strong&gt;&lt;div&gt;Q와 K는 내적 가능해야하므로 반드시 같은 차원이어야 한다. (&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)&lt;/div&gt;&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Q,K의 차원과 V의 차원(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d_v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)은 같을 필요가 없다&lt;/strong&gt;. V는 결국 상수배해서 가중평균 낼 것이기 때문이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;exp&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;mi&gt;exp&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;msub&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A(q,K,V) = \sum_i\frac{\exp(q\cdot k_i)}{\sum_j\exp(q\cdot k_j)}v_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.677227em;vertical-align:-0.667227em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.16195399999999993em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29971000000000003em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.01em&quot;&gt;&lt;span style=&quot;top:-2.655em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mop mtight&quot;&gt;&lt;span class=&quot;mop op-symbol small-op mtight&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.14964714285714287em&quot;&gt;&lt;span style=&quot;top:-2.1785614285714283em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.46032428571428574em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace mtight&quot; style=&quot;margin-right:0.19516666666666668em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop mtight&quot;&gt;&lt;span class=&quot;mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mtight&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.03148em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2818857142857143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.485em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mop mtight&quot;&gt;&lt;span class=&quot;mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mtight&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.03148em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.667227em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Attention 모듈의 input은 하나의 query 벡터, 모든 Key(를 concat한) 벡터와 모든 Value(를 concat한) 벡터가 된다.&lt;/li&gt;&lt;li&gt;분자는 i번째 key와 query 사이의 유사도, 즉 &lt;strong&gt;두 단어간의 유사도&lt;/strong&gt;가 된다. 여기에 해당 단어의 value 벡터를 곱한다.&lt;/li&gt;&lt;li&gt;분모는 모든 key에 대한 유사도의 총합이 된다.&lt;/li&gt;&lt;li&gt;따라서, (해당 단어의 유사도 / 전체 단어의 유사도 총합) 형태가 되므로 가중평균을 구성하게 된다. 이 때 출력 벡터는 value 벡터의 크기가 될 것이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;그렇다면 query를 여러개 쌓아 행렬 Q를 만든 뒤 한번에 표현해보도록 하자.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mtext&gt;softmax&lt;/mtext&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A(Q,K,V) = \text{softmax}(QK^T)V&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1413309999999999em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;softmax&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이를 행렬의 형태로 표현하면 다음 이미지와 같다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:23.046875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAFCAIAAADKYVtkAAAACXBIWXMAABYlAAAWJQFJUiTwAAABDklEQVQY01WLy0rDQBhG5wV8A/dufZHiE4hC30GEgiCk2i4MKKIIbnSjdNGVgquuhDYEDVRRMJCQBpvL5DKZDLlMJhknQkHP4ufw/RzAV1SUIndJSUKiAPpeiVGGwsBzY+jRFBUoJDiBrpuFfoEiRkuRACqoKmG4ak5fFsdTY6xDQtm5ah4+6zrKFSeVFOtItW1SfkIsTY3BzNCCvI1N01w6jrBFxdd696Cz37lTU87Xe9dgdzCJuaxaYGsPdE9mqH60MNiRwfbwbA7buCiKklJhKeMXOunPo4fvPBbxwS3oyk8+G5l443KyeaNoCbOzuv8WDd8jDbE25v9pfm9W86uPQHp1v0g70Iaz1esvP6C9/WeawqpTAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;qkv&quot; title=&quot;qkv&quot; src=&quot;/static/b89ecf5471f058093ce840999f4efaa5/2bef9/qkv.png&quot; srcSet=&quot;/static/b89ecf5471f058093ce840999f4efaa5/6f3f2/qkv.png 256w,/static/b89ecf5471f058093ce840999f4efaa5/01e7c/qkv.png 512w,/static/b89ecf5471f058093ce840999f4efaa5/2bef9/qkv.png 1024w,/static/b89ecf5471f058093ce840999f4efaa5/71c1d/qkv.png 1536w,/static/b89ecf5471f058093ce840999f4efaa5/b8471/qkv.png 2016w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\vert Q\vert&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 쿼리의 개수&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 쿼리의 차원&lt;/li&gt;&lt;li&gt;Q와 K의 &lt;strong&gt;내적을 위해서는 K가 transpose되어야 한다&lt;/strong&gt;. 내적의 결과로 나오는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;|Q|\times|K|&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 행렬에서 &lt;strong&gt;i번째 row는 i번째 query에 대한 input 벡터들의 유사도를 나타내는 row&lt;/strong&gt;가 된다. 이 연산이 끝나면 softmax를 이용하여 가중치 벡터로 변환된다.&lt;/li&gt;&lt;li&gt;가중치 벡터와 V 벡터와의 내적을 통해 나온 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;|Q|\times d_K&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 출력 행렬은 다시 기존의 Q와 동일한 형태를 이루며, &lt;strong&gt;출력 행렬의 i번째 row는 input Q의 i번째 row(query)에 대한 attention의 output&lt;/strong&gt;이 된다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;scaled-dot-product-attention&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#scaled-dot-product-attention&quot; aria-label=&quot;scaled dot product attention permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Scaled Dot-Product Attention&lt;/h3&gt;&lt;p&gt;위의 경우는 2차원 행렬을 가정하고 수행했지만, 실제로 연산을 수행할 때에는 q와 k가 &lt;strong&gt;n차원&lt;/strong&gt;일 수 있다.&lt;/p&gt;&lt;p&gt;이때, dimension이 커지면 커질수록 &lt;strong&gt;q와 k의 내적값의 쏠림현상&lt;/strong&gt;이 심해져서, attention이 효율적으로 학습되지 못한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이는 서로 독립인 random variable끼리 곱했을 때 분산이 1이되고, 이 곱들을 서로 더했을 때 분산이 고스란히 더해지기 때문이다. 내적 수행 과정에서는 차원이 클수록 random variable 값끼리 더 많이 곱하고 더해지게 되는데, 이 과정에서 분산이 너무 커진다. 이는 곧 내적값의 차이가 크게 나는 것을 의미한다.&lt;/li&gt;&lt;li&gt;이를 직관적으로 생각해보자. 2차원 평면상에 떨어져있던 여러 점들이, 3차원 공간으로 이동하면, 점들간의 거리는 어떻게 될까? 높이(z)라는 새로운 축이 생겼으므로, 당연히 점들간의 거리는 평면상보다 대체로 멀어지고, 최선의 경우(세 점들이 같은 평면에 있을 때)에야  동일한 거리를 가지게 된다.&lt;/li&gt;&lt;li&gt;따라서, 차원이 늘어난다는 것은 대체로 공간상의 점좌표로 표현되는 vector간의 거리가 sparse해 진다는 것, 즉 분산이 늘어난다는 것을 직관적으로 이해할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mtext&gt;softmax&lt;/mtext&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;msqrt&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/msqrt&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A(Q,K,V) = \text{softmax}\Bigg(\frac{QK^T}{\sqrt{d_k}}\Bigg)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.0000299999999998em;vertical-align:-1.25003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;softmax&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.5183309999999999em&quot;&gt;&lt;span style=&quot;top:-2.25278em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord sqrt&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.85722em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;padding-left:0.833em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.81722em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail&quot; style=&quot;min-width:0.853em;height:1.08em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;1.08em&quot; viewBox=&quot;0 0 400000 1080&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M95,702
c-2.7,0,-7.17,-2.7,-13.5,-8c-5.8,-5.3,-9.5,-10,-9.5,-14
c0,-2,0.3,-3.3,1,-4c1.3,-2.7,23.83,-20.7,67.5,-54
c44.2,-33.3,65.8,-50.3,66.5,-51c1.3,-1.3,3,-2,5,-2c4.7,0,8.7,3.3,12,10
s173,378,173,378c0.7,0,35.3,-71,104,-213c68.7,-142,137.5,-285,206.5,-429
c69,-144,104.5,-217.7,106.5,-221
l0 -0
c5.3,-9.3,12,-14,20,-14
H400000v40H845.2724
s-225.272,467,-225.272,467s-235,486,-235,486c-2.7,4.7,-9,7,-19,7
c-6,0,-10,-1,-12,-3s-194,-422,-194,-422s-65,47,-65,47z
M834 80h400000v40h-400000z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.18278000000000005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.93em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이런 문제를 해결하기 위해, 분산을 일정하게 유지하기 위한 방법으로 Q와 K의 내적을 &lt;strong&gt;차원(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)의 제곱근으로 나누어주는 테크닉&lt;/strong&gt;이 있다. 이 경우 분산이 분산의 제곱배만큼 축소되기 때문에, 차원이 몇차원이든 &lt;strong&gt;분산을 항상 1로 일정하게 Scaling&lt;/strong&gt;할 수 있다.&lt;/p&gt;&lt;h2 id=&quot;multi-head-attention&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#multi-head-attention&quot; aria-label=&quot;multi head attention permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Multi-Head Attention&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Multi-Head Attention&lt;/code&gt;&lt;/strong&gt;은 기존의 Attention 모듈을 좀 더 유용하게 확장한 모듈이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:222px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:102.25225225225225%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;mha&quot; title=&quot;mha&quot; src=&quot;/static/3899dc849f05e31eba1cc3c129936232/6bc44/mha.png&quot; srcSet=&quot;/static/3899dc849f05e31eba1cc3c129936232/6bc44/mha.png 222w&quot; sizes=&quot;(max-width: 222px) 100vw, 222px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mtext&gt;MultiHead&lt;/mtext&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;C&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;o&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;n&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;c&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;a&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;h&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;e&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;a&lt;/mi&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;d&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;h&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;e&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;a&lt;/mi&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;d&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;/msup&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;W&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;h&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;e&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;r&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;e&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;h&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;e&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;a&lt;/mi&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;d&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;A&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;t&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;t&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;e&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;n&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;t&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;i&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;o&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;msubsup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;/msubsup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;msubsup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;/msubsup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;msubsup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/msubsup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\text{MultiHead}(Q,K,V) = \mathrm{Concat(head_1,\dots,head_n)}W^O\\
\mathrm{Where\ head_i = Attention}(QW_i^Q,KW_i^K,VW_i^V)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord&quot;&gt;MultiHead&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1413309999999999em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathrm mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathrm mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.236103em;vertical-align:-0.276864em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31750199999999995em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathrm mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9592389999999998em&quot;&gt;&lt;span style=&quot;top:-2.4231360000000004em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.180908em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;Q&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.276864em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;Multi-head attention은 &lt;strong&gt;여러개의 attention 모듈을 동시에 사용&lt;/strong&gt;한다. 이 때 각 attention 모듈의 선형변환 파라미터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;(head)는 모듈마다 각각 다르다. 이 각기 다른 version의 모듈들을 이용하여 낸 output들을 &lt;strong&gt;concat&lt;/strong&gt;(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\times i)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;한 뒤, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^O&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8413309999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 선형변환하여 하나의 output을 만드는 형태이다.&lt;/p&gt;&lt;p&gt;이런 Multi-head attention을 사용하는 이유는, 동일한 입력문 기준으로도 &lt;strong&gt;필요에 따라 중점을 두어야 할 단어들이 다를 경우가 있기 때문&lt;/strong&gt;이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;가령, I am going to eat dinner라는 문장이 있다고 하자.&lt;/li&gt;&lt;li&gt;어떤 때는 &amp;#x27;내가&amp;#x27; 먹었다는 사실에 주목해야 해서 &amp;#x27;I&amp;#x27;에 집중해야 할 수도 있고, 어떤 때는 &amp;#x27;저녁&amp;#x27;을 먹었다는 사실에 주목하기 위해 &amp;#x27;dinner&amp;#x27;에 집중해야 할 수도 있다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;attention의-연산량&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#attention%EC%9D%98-%EC%97%B0%EC%82%B0%EB%9F%89&quot; aria-label=&quot;attention의 연산량 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;attention의 연산량&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : sequnce 길이&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : query / key 차원&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 합성곱 연산의 커널 사이즈&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;r&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : restricted self-attention의 이웃 사이즈&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;위와 같을 때, 기존 layer들의 연산량은 다음과 같다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:23.046875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAFCAIAAADKYVtkAAAACXBIWXMAABYlAAAWJQFJUiTwAAAAvUlEQVQY0z2QyQqEMBBE8/8fpwiC4sENNVEEd+Ma541hpg+hulLV1YlY17Usy6qqlmUBa62bponjGLIoCqUUIAgCKWWe52ma2hP+eR6xbZtVTNN0Xdd5nvu+d11X13Xf92CuXNeFQTmOY9u2NuZrpvd9P4oizPTGGE6keNCB8WRZZnnGkUyAVQo8nuc5jsOq6i1Y+Rbh932TkyQJBjCvC8NwnmfzljiOg4YRrMDgYRh4Nlv9v8Bi/StLsg7hH8THGKWUjEPkAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;calc&quot; title=&quot;calc&quot; src=&quot;/static/467f778298e65913d0aaf45696756b2f/2bef9/calc.png&quot; srcSet=&quot;/static/467f778298e65913d0aaf45696756b2f/6f3f2/calc.png 256w,/static/467f778298e65913d0aaf45696756b2f/01e7c/calc.png 512w,/static/467f778298e65913d0aaf45696756b2f/2bef9/calc.png 1024w,/static/467f778298e65913d0aaf45696756b2f/71c1d/calc.png 1536w,/static/467f778298e65913d0aaf45696756b2f/a878e/calc.png 2048w,/static/467f778298e65913d0aaf45696756b2f/437e0/calc.png 2094w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Complexity per Layer&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;Total Computational Complexity per Layer를 의미하는 것으로, 총 연산양의 복잡도를 의미한다. 논문 원문에서는 (연산양에 따른) 시간복잡도를 의미한 듯 한데, 공간복잡도로 해석해도 큰 무리는 없는 듯하다. 연산을 처리하는 디바이스가 1개인 상황을 가정한다고 생각하면 얼추 들어맞는다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Self-attention&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Q \times K^T = (n \times d) \times (d \times n)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8413309999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;Q와 K를 내적하므로 각 길이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 제곱하고, 이를 모든 dimension마다 계산해야 하므로 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 곱해야한다.  - &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O(n^2\cdot d)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.064108em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Recurrent&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;time step의 개수가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이고, 매 time step마다 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(d\times d)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 크기의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_{hh}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 곱한다. - &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O(n\cdot d^2)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.064108em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_{hh}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 dimension &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 hidden state vector의 크기로, 하이퍼파라미터이므로 직접 정해줄 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Sequential Operations&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;해당 연산을 얼마의 시간내에 끝낼 수 있는가를 나타낸 것으로, 연산양 자체는 무한히 많은 GPU의 병렬연산으로 한번에 처리할 수 있음을 가정한다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Self-attention&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;시퀀스의 길이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 길어질수록 지수배로 연산복잡도가 늘어난다.&lt;/li&gt;&lt;li&gt;이는 모든 Q와 K의 내적값을 모두 저장하고 있어야하기 때문이다.따라서 &lt;strong&gt;일반적인 Recurrent보다 훨씬 많은 메모리를 필요&lt;/strong&gt;로 하게 된다.&lt;/li&gt;&lt;li&gt;그러나 GPU는 이런 형태의 행렬 연산 병렬화에 특화되어 있고, 따라서 무한히 많은 GPU를 가지고만 있기만 하다면 이를 병렬화하여 계산할 수 있으므로, 시간 복잡도는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O(1)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Recurrent&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;이전 time step의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_{t-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.902771em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 제공되어야 그것을 input으로 다음 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 계산할 수 있기 때문에, 불가피하게 시간 복잡도는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O(n)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Maximum Path Length&lt;/code&gt;&lt;/strong&gt; - 두 단어 간의 경로 거리&lt;ul&gt;&lt;li&gt;&lt;strong&gt;Self-attention&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;두 단어간의 유사도를 구할 때, 행렬 연산으로 바로 곱할 수 있으므로 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O(1)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Recurrent&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;어떤 단어 a가 어느정도 떨어진 단어 b에 도달하기까지 recurrent cell을 하나씩 통과해야하기 때문에 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O(n)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;block-based-model&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#block-based-model&quot; aria-label=&quot;block based model permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Block-Based Model&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:798px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:180.078125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;transformer-encoder&quot; title=&quot;transformer-encoder&quot; src=&quot;/static/d0dc87f285a8bdd8c7c39e39c88d109a/898f6/transformer-encoder.png&quot; srcSet=&quot;/static/d0dc87f285a8bdd8c7c39e39c88d109a/6f3f2/transformer-encoder.png 256w,/static/d0dc87f285a8bdd8c7c39e39c88d109a/01e7c/transformer-encoder.png 512w,/static/d0dc87f285a8bdd8c7c39e39c88d109a/898f6/transformer-encoder.png 798w&quot; sizes=&quot;(max-width: 798px) 100vw, 798px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;처음에 시작해 Multi-Head Attention으로 가는 세개의 화살표는 각각 Q,K,V를 의미한다. 각 head마다 들어가게 된다.&lt;/p&gt;&lt;p&gt;그 연산 이후에 진행되는 &lt;strong&gt;Add&amp;amp;Norm&lt;/strong&gt; 구간은 무엇일까?&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Add - &lt;strong&gt;&lt;code&gt;Residual Connection&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;CV 쪽에서 깊은 레이어를 만들 때 graident vanishing을 해결하면서 더 깊은 층을 쌓도록 하는 효과적인 모델이다.&lt;/li&gt;&lt;li&gt;주어진 input vector를 Multi-Head Attention의 encoding output에 그대로 더하여 새로운 output을 만들어서, 학습시에 Multi-Head Attention이 &lt;strong&gt;입력 벡터 대비 정답 벡터와의 &amp;#x27;차이나는 정보&amp;#x27;만 학습&lt;/strong&gt;하도록 할 수 있다.&lt;/li&gt;&lt;li&gt;이 때, &lt;strong&gt;Multi-Head Attention output과 input 벡터의 크기가 완전히 동일하도록 유지&lt;/strong&gt;해야 더할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Norm - &lt;strong&gt;&lt;code&gt;Normalization&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;일반적으로 신경망에서 사용되는 normalization은, (평균,분산)을 (0,1)로 만든 뒤, 원하는 평균과 분산을 주입할 수 있도록 하는 &lt;code&gt;선형변환(affine transformation)&lt;/code&gt;으로 이루어진다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;Batch Norm&lt;/code&gt;&lt;ul&gt;&lt;li&gt;각 원소에 평균을 빼고, 표준편차로 나눈다. → (평균,분산)==(0,1)&lt;/li&gt;&lt;li&gt;affine transforamtion하여 원하는 평균과 분산으로 만든다.&lt;ul&gt;&lt;li&gt;ex ) &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mstyle&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mstyle mathcolor=&quot;Green&quot;&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/mstyle&gt;&lt;mo&gt;→&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mtext&gt;평균&lt;/mtext&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mtext&gt;분산&lt;/mtext&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mstyle mathcolor=&quot;Green&quot;&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/mstyle&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mstyle&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y= \textcolor{red}2x +\textcolor{Green}3 \rarr (평균,분산) = (\textcolor{Green}3,\textcolor{red}2)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:Green&quot;&gt;3&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;→&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;평&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;균&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;분&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;산&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:Green&quot;&gt;3&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;이 때 2와 3은 Optimization 과정에서 최적화 대상인 파라미터가 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;Layer Norm&lt;/code&gt;&lt;ul&gt;&lt;li&gt;Batch Norm과 마찬가지 방법으로 수행하되, 여러 layer가 붙어있는 행렬을 대상으로, 한 layer마다 수행한다.&lt;/li&gt;&lt;li&gt;affine transformation은 각 layer의 동일한 node 기준으로 수행한다.(normalization이 column 단위였다면 affine transformation은 row 별)&lt;/li&gt;&lt;li&gt;Batch Norm과는 일부 차이점이 있지만, 큰 틀에서 &lt;strong&gt;학습을 안정화&lt;/strong&gt;한다는 점은 동일하다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Add&amp;amp;Norm 구간을 거치고 나온 output은 다시 fully connected layer(Feed Forward)에 통과시켜 Word의 인코딩 벡터를 변환한다. 이후 다시 Add&amp;amp;Norm을 한번 더 수행하는 것까지를 끝으로 Transformer의 (self-attention 모듈을 포함한) &lt;strong&gt;&lt;code&gt;Block Based Model&lt;/code&gt;&lt;/strong&gt;이 완성된다.&lt;/p&gt;&lt;h3 id=&quot;positional-encoding&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#positional-encoding&quot; aria-label=&quot;positional encoding permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Positional Encoding&lt;/h3&gt;&lt;p&gt;RNN과 달리 self-attention 모듈 기반의 Block Based Model로 인코딩하는 경우, 순서를 고려하지 않기 때문에 input 단어들의 순서가 바뀌어도 output은 동일하다. 이는 K와 Q간의 유사도를 구하고 V로 가중치를 구해 가중합(이때 softmax를 통과한 값이므로 가중합 자체가 가중평균이다)을 도출하는 과정에서, sequence를 고려하지 않기 때문이다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Positional Encoding&lt;/code&gt;&lt;/strong&gt;은 이를 해결하기위해 벡터 내의 특정 원소에 해당 word의 순서를 알아볼 수 있는 , 마치 &lt;strong&gt;&lt;div&gt;지문과도 같은 unique한 값을 추가하여 sequence를 고려하는 것&lt;/div&gt;&lt;/strong&gt;을 말한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때, unique한 값은 여러 주기함수의 출력 함수값을 합쳐 사용한다. 주기함수는 입력값 x의 위치에 따라 출력값이 변하기 때문이다.&lt;/li&gt;&lt;li&gt;단, 하나의 주기함수만 사용하면 동일한 함수값을 가지는 구간이 생기므로, &lt;strong&gt;서로 다른 여러 주기함수의 출력값들을 모두 합쳐서 사용&lt;/strong&gt;한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이렇게 특수한 값을 추가하여 인코딩하게 되면, input 단어의 순서가 바뀌었을 때 output 값도 달라지게 되어 순서를 구별할 수 있는 모델이 된다.&lt;/p&gt;&lt;h3 id=&quot;learning-rate-scheduler&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#learning-rate-scheduler&quot; aria-label=&quot;learning rate scheduler permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Learning Rate Scheduler&lt;/h3&gt;&lt;p&gt;기존의 모델에서 학습률(learning rate)는 하이퍼파라미터로, 학습 내내 고정되어있는 값이었다. 그러나 학습의 과정동안 효율적인 학습률은 계속 바뀌기 마련이므로, 이를 학습 과정 내에서 효과적으로 바꾸어 줄 수 있는 방식으로 &lt;strong&gt;&lt;code&gt;Learning Rate Scheduler&lt;/code&gt;&lt;/strong&gt;가 나오게 되었다.&lt;/p&gt;&lt;h2 id=&quot;디코더-구조&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%94%94%EC%BD%94%EB%8D%94-%EA%B5%AC%EC%A1%B0&quot; aria-label=&quot;디코더 구조 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;디코더 구조&lt;/h2&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:616px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:257.8125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAA0CAIAAAAFR49sAAAACXBIWXMAABYlAAAWJQFJUiTwAAAGBUlEQVRIx51WCVfaWBT290877ahtp1WrKAqiLIEQliyEJQHCviQkhKyyKy64VKwLc4UZOzMHpDP3hHNewvveve9+333vLo3/Zk9PT+P/Yksvo8fHx/Zfdnx8fHd3t3C5P8H39/ffv38fDAbNZlNV1VarBa+LwdO/2522IAj1er1Wq8GgOjFY6HX8j7CfHp8ndTtdTdWmsTw8PCzeM6ytaAov8qquMikWpwijadRkUZSl13f+DB5eDNdcmwdJ5O32e3g+Ob/8YnnnYDxfvJuark1zORcMEVbEakmq5PkCESP8BJYt50piuSAUr6+vF3ieOj/un9xc3dSleiFfHN3ewevZ6dk0Ea/teTi8sO37o4zg8cV29tCtHbf9MBxP1uwHYV03FoQNlOYL1SovF0s1nIh6ECyXr/JCI5srX15eLgi73+/zfK1S4RuySpG0z+tXFb1cqgqCKMuNUqkMmpm5xBJki47EUP8O4ltDfBtO9+/7Bx+8vg1/YNPr/0pRtK410+nMzc2MzC3BV5blotFtP/rWi/waCq0EAr8hnrcE8YEgPpIkleEKDJOEvMwE32S5nBUJbvrwr57g243dd1v2DRdmQcmPNnevOzgy26EQPpOzpdvRbYyOHXICMxwTzaE9K1mZEqb0U1fjg5LCC7KhN3GcnAO+nYAzfGxwb8+Ih0XFnhW3mTLeGjpKSpWvvwaGsLl0xkMmnDHuMMJaUNziC7volC9ZOCAT3c7ANFrheWFfXV0BSeeDq6bePTu5zHGlHYstnynHKIaNc5KkpFNZlk3NJHzJMAyOy0fjTiq6S9F7OLXjcC67PB9Q7LPb+8mLeE2jLfDi2dnZDLAkiZlMIYBvO5HlA/eKE1lx+1adyGoAX8fC65EIzaXziQQ727MkSblcKezc9FvfI9vvXV/fuDZ+8VmXA7ZV1LrcbfeOjrpzqWrIcjyRqqcc54Jbojf7xcPj0mE9utkrHjTz+/WaZBhtHCdmg2W5Ho0yVXp73AwVsA8C+UWMrAvE51vV38zuSYJomp25YFGEE09iSDTs2srHA2H3DurYyET9pNfqta/1useaalJkBA5jRVGAmr8vsYRhGKifpOL7Bx5FbUZoBqpSVdskmVC152KqSyqIBBJ2cXHx7du3f3iGJWk6dmT2dK1jGr1ySbTbXQQRi0VTxWKtVKymkplUihsOh7MVBtlmk34qYovQjmDQsu9Y8WOb/sBXt+fzszyGN3N5hkgYJh0it1zedy7kvS+wigY/un3LYWoNHhRFQUJQz3M9c1wu6LRQ7rXDrd/21984Nt54d1cJ5xpiXe20YDvNuTyPRrcROiGx9vFxRGN3apF1ldluxLfGHbKVt4t8bcLz/JKko0yddZjp3TK+bqRsZtpWIdZVdtdI7TUk+TXwCMB0op7cB1dyzKIlre3svpKw3MhoO++oi5L5isKeE8amE/4tJWlLYxuEYyXm+RRHPgvRXdb7u6boht6a6xlEA0xIkmrZ3hNFJRjEXS5E05puD5pg0k+P44asvwK+LBTK46dx86ijKkYqmcX8wUpZgGdwcn46GFYqNZDn7JKE9wyX5atwvmum2QLarFYbCBbowbAgimLhEAHn1MPD/Qww/CDh3YlBW8HzfLFYBPJhDF/Oz8/hPoEJizqDyaq9Xk/Tnu9k6C86nQ6kE77oug69BvQqSkORZflHVb0gp59gEkEQHMcpjQZ4hi4Fiul0YhALRAGx/Nvz3f39zSS2hqLkC4WHpydV1wulUq/fbzVbEAvcdVDP4FbX9H/ueXQLRVxJMGwoTLg9YaeLI6l8JEr70LPT0+n9DDaNbtql/ACfnZ8Lsfi42+8WSiqbEqMxPZUe9wfNbN7U9UmjNP9yB3CZiqgMWyWpMoSAkzksUI/G4TENYwEYIknRdJqiuAiNI97AoTNNRWKBYDwUhrZkcUPzMgMSAyS90ofMaOJeqAI+WpOu8SVDP9X1jkajk5MToBFkDCfWVFWLwdMZMBsuPe0vm4rhpzyDBkF9oCfTNGEM/W6j0YD9/xR4qkHwBsf49dU14DvdDsT/s3v+H/YH1/fjy94WfNIAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;transformer-decoder&quot; title=&quot;transformer-decoder&quot; src=&quot;/static/ba409d1cf7c2c83aafc61a4a293dd1cb/40040/transformer-decoder.png&quot; srcSet=&quot;/static/ba409d1cf7c2c83aafc61a4a293dd1cb/6f3f2/transformer-decoder.png 256w,/static/ba409d1cf7c2c83aafc61a4a293dd1cb/01e7c/transformer-decoder.png 512w,/static/ba409d1cf7c2c83aafc61a4a293dd1cb/40040/transformer-decoder.png 616w&quot; sizes=&quot;(max-width: 616px) 100vw, 616px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;Outputs가 디코더의 입력으로 들어올 때, 기존의 ground truth 문장에서 앞쪽에는 &lt;code&gt;&amp;lt;SOS&amp;gt;&lt;/code&gt; 토큰을 붙여 들어오므로, 한칸 밀린(shfited right) 형태로 들어오게 된다.&lt;/p&gt;&lt;p&gt;디코더에서 Attention 모듈을 포함한 한차례의 과정을 거친 후 다음 Multi-Head Attention으로 갈 때, 디코더의 hidden state vector를 입력 Q로 넘겨준다. 그런데, 나머지 K와 V 입력은 외부, 즉 인코더의 최종 출력으로부터 온다. 즉, 이 부분은 디코더의 hidden state vector를  기준, 즉 Q로 해서 인코더의 hidden state vector K, V 를 가중하여 가져오는, &lt;strong&gt;인코더와 디코더간의 Attention 모듈&lt;/strong&gt;이 된다.&lt;/p&gt;&lt;p&gt;이 후 이미지에 나온 대로의 연산을 거치다가, 디코더의 최종 output 값이 Linear Layer와 Softmax를 거쳐 확률분포의 형태로 출력된다. 이 값은 Softmax-with-loss 손실함수를 통해 학습된다.&lt;/p&gt;&lt;h3 id=&quot;masked-self-attention&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#masked-self-attention&quot; aria-label=&quot;masked self attention permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Masked Self-Attention&lt;/h3&gt;&lt;p&gt;Self-Attention 모델에서, 임의의 단어 a는 Q와 K의 내적을 통해 자신과 모든 단어들의 관계를 다 알수 있다. 이 때, 학습 당시에는 배치 프로세싱을 위해 a 뒤의 단어들까지 모두 고려하도록 학습이 진행되나, 사실 &lt;strong&gt;실제 디코딩 상황을 고려한다면 a 뒤의 단어를 알아서는 안된다&lt;/strong&gt;. 이는 뒤의 단어를 추론해야 하는 상황에서 뒤에 어떤 단어가 있는지 미리 알고있는, 일종의 cheating 상황이기 때문이다. 이러면 당연히 학습은 제대로 되지 않게 된다.&lt;/p&gt;&lt;p&gt;디코더 과정의 이미지 중 &lt;strong&gt;&lt;code&gt;Masked Self-attention&lt;/code&gt;&lt;/strong&gt;은 이를 해결하기 위한 방법으로, 기존의 attention 모듈에서 Q, K 내적과 softmax를 통과한 값에서 현재 단어 a의 뒤에 있는 단어들을 key 값으로 계산된 셀들을 모두 삭제한다. &lt;code&gt;Mask&lt;/code&gt; 라는 단어는 이처럼 &lt;strong&gt;뒤쪽의 정보를 가린다(mask)&lt;/strong&gt;는 의미다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:68.359375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;masked-self-attention&quot; title=&quot;masked-self-attention&quot; src=&quot;/static/dd65ab455be26cc1d2b62a11095fdda4/2bef9/masked-self-attention.png&quot; srcSet=&quot;/static/dd65ab455be26cc1d2b62a11095fdda4/6f3f2/masked-self-attention.png 256w,/static/dd65ab455be26cc1d2b62a11095fdda4/01e7c/masked-self-attention.png 512w,/static/dd65ab455be26cc1d2b62a11095fdda4/2bef9/masked-self-attention.png 1024w,/static/dd65ab455be26cc1d2b62a11095fdda4/71c1d/masked-self-attention.png 1536w,/static/dd65ab455be26cc1d2b62a11095fdda4/e92cd/masked-self-attention.png 1778w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위의 이미지는 [I go home → 나는 집에 간다] 라는 번역을 수행하는 사례인데, Q,K의 내적을 통해 얻은 정사각 행렬을 표현하고 있다. 이 때 주대각선 위쪽의 값들은 query보다 key가 뒤쪽의 단어들인 경우로, 이 셀들의 정보를 그대로 둔 채로 학습시키지 못하도록 해당 값들을 0으로 대체한다. 그 이후, 남은 주대각선 이하의 셀들만 가지고, row단위로 총합이 1이 되도록 normalize 한 정보를 최종 output으로 내보낸다.&lt;/p&gt;&lt;hr/&gt;&lt;h3 id=&quot;reference&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#reference&quot; aria-label=&quot;reference permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Reference&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;http://jalammar.github.io/illustrated-transformer/&quot;&gt;The Illustrated Transformer&lt;/a&gt; - &lt;a href=&quot;https://nlpinkorean.github.io/illustrated-transformer/&quot;&gt;한국번역&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://openaccess.thecvf.com/content_ECCV_2018/papers/Yuxin_Wu_Group_Normalization_ECCV_2018_paper.pdf&quot;&gt;Group Normalization&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://lv99.tistory.com/26&quot;&gt;Transformer: All you need is Attention (설명/요약/정리)&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://jeongukjae.github.io/posts/attention-is-all-you-need-review/&quot;&gt;📃 Review of &amp;quot;Attention Is All You Need&amp;quot;&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[NLP 05 - Seq2Seq와 Attention]]></title><description><![CDATA[Sequence to Sequence with attention by 주재걸 교수님, BoostCamp AI Tech 4주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/26_seq2seq_with_attention/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/26_seq2seq_with_attention/</guid><pubDate>Wed, 17 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;seq2seq&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#seq2seq&quot; aria-label=&quot;seq2seq permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Seq2Seq&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:30.078125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAAAsSAAALEgHS3X78AAABAUlEQVQY0z2QW2/CMAyF+/9/1F7QNGkPk/bQbmEI2oaW0txokrZxbsyA4JMi2efYUo6L6xMAGO7knF9iyjdiSlhba7UxN/FJ8ZrL6RoghhAfHb64TMEIJdUyWyWE5NxM00VJWBcsoveFdmKca247vlJhu57/NV3FDBXxfN6+0/KzNr5hspG6ZopK3Uj0poPQaoFCzn19+R7MgU5Vr3flbvO7/zgq0i6kq952XxtyAdKPZJTbUf2cGOGanFg5CGZdgf/U2qSUMVKKyYFf19UYM3Jp2dHJ3kHwHkIIGNID5JQjhgsBc+FycM5575VSbdtSSnETWxTT/U4PcAxviS5aL/EfHtdUbC8OfYoAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;seq2seq-diagram&quot; title=&quot;seq2seq-diagram&quot; src=&quot;/static/f0e6589f9c8c025e217dd7d3c12c0ac9/2bef9/seq2seq-diagram.png&quot; srcSet=&quot;/static/f0e6589f9c8c025e217dd7d3c12c0ac9/6f3f2/seq2seq-diagram.png 256w,/static/f0e6589f9c8c025e217dd7d3c12c0ac9/01e7c/seq2seq-diagram.png 512w,/static/f0e6589f9c8c025e217dd7d3c12c0ac9/2bef9/seq2seq-diagram.png 1024w,/static/f0e6589f9c8c025e217dd7d3c12c0ac9/71c1d/seq2seq-diagram.png 1536w,/static/f0e6589f9c8c025e217dd7d3c12c0ac9/bd82c/seq2seq-diagram.png 1585w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;RNN의 Many-to-Many 형태에 해당한다.&lt;/p&gt;&lt;p&gt;인코더(Encoder)와 디코더(Decoder)로 구성되어 있으며, 인코더와 디코더는 별개의 파라미터를 가진 서로 다른 RNN 모델(이미지에서는 LSTM)이다.&lt;/p&gt;&lt;p&gt;인코더의 마지막 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;은 디코더의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_0&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 되며, 디코더는 첫 입력으로 &lt;code&gt;&amp;lt;SOS&amp;gt;&lt;/code&gt; 토큰을 받고, 마지막 출력으로 &lt;code&gt;&amp;lt;EOS&amp;gt;&lt;/code&gt;토큰을 낸 뒤 생성을 마친다.&lt;/p&gt;&lt;p&gt;따라서, 인코더의 모든 정보들은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_0&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이라는 하나의 hidden state vector에 담겨서 디코더에 반영된다.&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;seq2seq-with-attention&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#seq2seq-with-attention&quot; aria-label=&quot;seq2seq with attention permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Seq2Seq with attention&lt;/h2&gt;&lt;h3 id=&quot;기존-seq2seq의-한계&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B8%B0%EC%A1%B4-seq2seq%EC%9D%98-%ED%95%9C%EA%B3%84&quot; aria-label=&quot;기존 seq2seq의 한계 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;기존 Seq2Seq의 한계&lt;/h3&gt;&lt;p&gt;기존의 Seq2Seq 모델은 RNN의 특성상 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 dimension이 정해져 있어서, 아무리 시퀀스 데이터의 단어가 많아지더라도 그 정보들을 모두 압축하여 같은 크기의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 욱여넣어야했다.&lt;/p&gt;&lt;p&gt;또, LSTM에서 아무리 Long-term Dependency 문제를 해결했다고 하더라도, 마지막 time step을 기준으로 할 때 훨씬 이전에 있는 time step의 정보들을 모델을 거치면서 점점 변질되거나 소실되어 정보를 잘 저장하지 못하는 문제가 발생했다.&lt;/p&gt;&lt;p&gt;이에 대한 대안으로 입력값의 순서를 바꾸는 경우(i.e. I go home → go I home)도 연구되곤 했다.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;attention-module-적용&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#attention-module-%EC%A0%81%EC%9A%A9&quot; aria-label=&quot;attention module 적용 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Attention Module 적용&lt;/h3&gt;&lt;p&gt;&lt;img src=&quot;https://3.bp.blogspot.com/-3Pbj_dvt0Vo/V-qe-Nl6P5I/AAAAAAAABQc/z0_6WtVWtvARtMk0i9_AtLeyyGyV6AI4wCLcB/s1600/nmt-model-fast.gif&quot; alt=&quot;seq2seq-with-attention&quot;/&gt;&lt;/p&gt;&lt;p&gt;이와 달리, attention 모듈을 활용한 Seq2Seq 모델은, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_0&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 하나에 모든 인코더의 hidden state vector을 반영하는것이 아니라, 각 time step의 모든 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 전부 디코더에 제공한다. 디코더는 단어를 하나씩 생성하며, 그 때 그 때 필요한 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;들을 선별하여 사용하게 된다.&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;과정&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B3%BC%EC%A0%95&quot; aria-label=&quot;과정 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;과정&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:60.546875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;attention-translation&quot; title=&quot;attention-translation&quot; src=&quot;/static/bd7b75416767980d5c86f5a96a208530/2bef9/attention-translation.png&quot; srcSet=&quot;/static/bd7b75416767980d5c86f5a96a208530/6f3f2/attention-translation.png 256w,/static/bd7b75416767980d5c86f5a96a208530/01e7c/attention-translation.png 512w,/static/bd7b75416767980d5c86f5a96a208530/2bef9/attention-translation.png 1024w,/static/bd7b75416767980d5c86f5a96a208530/387f2/attention-translation.png 1518w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위의 이미지의 과정을 풀어보자.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;프랑스어를 인코더에 집어넣고, 마지막으로 나오는 출력값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;4&lt;/mn&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_4^{(e)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.311108em;vertical-align:-0.26630799999999993em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.433692em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.26630799999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 디코더가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_0&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 받는다.&lt;/li&gt;&lt;li&gt;디코더는 입력값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 Start Token에 해당하는 워드 임베딩 벡터를 받는다.&lt;/li&gt;&lt;li&gt;디코더는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;과 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_0&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 이용하여 디코더의 hidden state vector &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_1^{(d)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.311108em;vertical-align:-0.26630799999999993em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.433692em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.26630799999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 얻는다.&lt;/li&gt;&lt;li&gt;인코더의 각 time step에서 나온 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;4&lt;/mn&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_1^{(e)},h_2^{(e)},h_3^{(e)},h_4^{(e)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.311108em;vertical-align:-0.26630799999999993em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.433692em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.26630799999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.433692em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.26630799999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.433692em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.26630799999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.433692em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.26630799999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 각각에 대해 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_1^{(d)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.311108em;vertical-align:-0.26630799999999993em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.433692em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.26630799999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 내적하여, 다음 단어 예측에 사용할 Attention Score를 매긴다.&lt;/li&gt;&lt;li&gt;Attention Score를 logit 벡터로 생각하고 softmax 함수를 적용하여, 각각의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t^{(e)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.2905559999999998em;vertical-align:-0.24575599999999992em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.454244em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2197999999999998em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.24575599999999992em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대응되는 확률값(Attention Distribution)을 뽑아낸다.&lt;ul&gt;&lt;li&gt;이 때, Attention Distribution은 확률값이므로 합이 1이 되는데,  이렇게 합이 1인 형태의 가중치를 Attention Module 내에서는 &lt;strong&gt;&lt;code&gt;Attention Vector&lt;/code&gt;&lt;/strong&gt;라고 부른다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;해당 확률값들은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t^{(e)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.2905559999999998em;vertical-align:-0.24575599999999992em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.454244em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2197999999999998em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.24575599999999992em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 적용될 가중치로 사용할 수 있다. 이미지에서는, 다음 단어 생성에 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_1^{(e)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.311108em;vertical-align:-0.26630799999999993em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.433692em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.26630799999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 가장 많이 사용될 것이라는 것을 보여주고 있다. 각 가중치와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t^{(e)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.2905559999999998em;vertical-align:-0.24575599999999992em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.454244em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2197999999999998em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.24575599999999992em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 곱한 뒤 합산하여 가중평균(Attention Output, &lt;strong&gt;&lt;code&gt;context vector&lt;/code&gt;&lt;/strong&gt;)을 낸다.&lt;ul&gt;&lt;li&gt;이 때, 위의 이미지에서 Attention scores, Attention distribution이 있는 부분들을 묶어 &lt;strong&gt;&lt;code&gt;Attention Module&lt;/code&gt;&lt;/strong&gt;로 말할 수 있으며, Attention 모듈의 입력은 각각의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t^{(e)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.2905559999999998em;vertical-align:-0.24575599999999992em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.454244em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2197999999999998em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.24575599999999992em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_1^{(d)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.311108em;vertical-align:-0.26630799999999993em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.433692em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.26630799999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 되고, 출력은 Attention output이 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Attention output과 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_1^{(d)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.311108em;vertical-align:-0.26630799999999993em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.433692em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.26630799999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 concat되어서 output layer의 입력으로 들어가고, output layer는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\hat{y_1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 값, 즉 &amp;#x27;the&amp;#x27;라는 글자를 내보낸다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;이 과정을 반복하면 아래와 같은 이미지가 된다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:60.9375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;attention-translation2&quot; title=&quot;attention-translation2&quot; src=&quot;/static/d2475b8928e42e8315b93a837857476a/2bef9/attention-translation2.png&quot; srcSet=&quot;/static/d2475b8928e42e8315b93a837857476a/6f3f2/attention-translation2.png 256w,/static/d2475b8928e42e8315b93a837857476a/01e7c/attention-translation2.png 512w,/static/d2475b8928e42e8315b93a837857476a/2bef9/attention-translation2.png 1024w,/static/d2475b8928e42e8315b93a837857476a/71c1d/attention-translation2.png 1536w,/static/d2475b8928e42e8315b93a837857476a/c4842/attention-translation2.png 1634w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;결국, 디코더의 hidden state vector &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t^{(d)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.2905559999999998em;vertical-align:-0.24575599999999992em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.454244em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2197999999999998em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.24575599999999992em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 다음과 같은 두가지 역할을 한다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\hat{y_t}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 도출하는 output layer의 input중 하나&lt;/li&gt;&lt;li&gt;인코더의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;[h_1^{(e)},\dots,h_n^{(e)}]&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.311108em;vertical-align:-0.26630799999999993em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.433692em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.26630799999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.5834080000000004em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.11659199999999997em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 각각 얼마나 반영하여 Attention output을 낼 지 (가중치를) 결정&lt;/li&gt;&lt;/ol&gt;&lt;h3 id=&quot;backpropagation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#backpropagation&quot; aria-label=&quot;backpropagation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Backpropagation&lt;/h3&gt;&lt;p&gt;Seq2Seq with attention의 역전파 과정에서는, Encoder 단에서 잘못 가져온 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;[h_1^{(e)},\dots,h_n^{(e)}]&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.311108em;vertical-align:-0.26630799999999993em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.433692em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.26630799999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.5834080000000004em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.11659199999999997em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 가중치들을 다시 재조정하여, 원하는 정보가 잘 선택 될 수 있도록 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t^{(d)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.2905559999999998em;vertical-align:-0.24575599999999992em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.454244em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2197999999999998em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.24575599999999992em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 학습한다.&lt;/p&gt;&lt;h3 id=&quot;teacher-forcing&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#teacher-forcing&quot; aria-label=&quot;teacher forcing permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Teacher Forcing&lt;/h3&gt;&lt;p&gt;처음에 모델이 작동할 때, 기존에 주어진 파라미터 정보가 얼마 없으므로 모델이 잘못된 예측을 할 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;i.e. I want to go home → go to home ....&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;학습과정에서는 모델이 잘못된 예측을 하더라도 다음 input에 ground truth인 정답 input값을 집어넣어주어 파라미터를 학습시킬 수 있다.  이를 &lt;strong&gt;&lt;code&gt;Teacher Forcing&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;/p&gt;&lt;p&gt;그러나 실제 inference 과정에서는 ground truth를 줄 수 없으므로, 학습 환경과 실제 추론 환경이 다르게 되어 파라미터가 제대로 최적화되지 않는다. 이를 피하기 위해 일정 부분까지는 파라미터를 teacher forcing으로 학습시키다가, 이후에는 완전한 inference로 학습하기도 한다.&lt;/p&gt;&lt;h3 id=&quot;attention-scores-메커니즘&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#attention-scores-%EB%A9%94%EC%BB%A4%EB%8B%88%EC%A6%98&quot; aria-label=&quot;attention scores 메커니즘 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Attention Scores 메커니즘&lt;/h3&gt;&lt;p&gt;위에서 Attention Score는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t^{(e)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.2905559999999998em;vertical-align:-0.24575599999999992em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.454244em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2197999999999998em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.24575599999999992em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t^{(d)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.2905559999999998em;vertical-align:-0.24575599999999992em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.454244em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2197999999999998em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.24575599999999992em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 유사도 역할을 하며, 이를 구하기 위해 내적을 했었다.&lt;/p&gt;&lt;p&gt;단순히 내적(dot product)을 하는 방식 외에도 Attention Scores를 구하는 다양한 방식들이 제안되었다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:47.265625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;attention_scores&quot; title=&quot;attention_scores&quot; src=&quot;/static/0c1b956e236b2042bc48ecfe6149adbf/2bef9/attention_scores.png&quot; srcSet=&quot;/static/0c1b956e236b2042bc48ecfe6149adbf/6f3f2/attention_scores.png 256w,/static/0c1b956e236b2042bc48ecfe6149adbf/01e7c/attention_scores.png 512w,/static/0c1b956e236b2042bc48ecfe6149adbf/2bef9/attention_scores.png 1024w,/static/0c1b956e236b2042bc48ecfe6149adbf/71c1d/attention_scores.png 1536w,/static/0c1b956e236b2042bc48ecfe6149adbf/a878e/attention_scores.png 2048w,/static/0c1b956e236b2042bc48ecfe6149adbf/8126d/attention_scores.png 2316w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;generalized dot product&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;디코더의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 인코더 중 한 단어의 hidden state vector &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mo&gt;ˉ&lt;/mo&gt;&lt;/mover&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\bar{h}_s&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.9812199999999999em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8312199999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.26344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;ˉ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 내적하는 과정에서, 중간에 하나의 가중치행렬(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_a)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 추가한다.&lt;/li&gt;&lt;li&gt;정사각행렬 내부의 원소들을 새로운 파라미터로 두어, Attention scores를 구하는 방법을 학습시킬 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;concat&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;디코더의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 인코더 중 한 단어의 hidden state vector &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mo&gt;ˉ&lt;/mo&gt;&lt;/mover&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\bar{h}_s&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.9812199999999999em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8312199999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.26344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;ˉ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 concat(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo separator=&quot;true&quot;&gt;;&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;;)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;하여 입력값으로 주고, scalar 값을 도출하는 새로운 신경망을 추가한다.&lt;/li&gt;&lt;li&gt;이 경우 가중치 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_a&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 곱한 뒤 비선형 변환 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;tanh&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tanh&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;tanh&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 거치고, 마지막으로 선형변환인 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v_a^T&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.088331em;vertical-align:-0.247em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 거쳐 스칼라 값(score)을 출력한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;attention의-의의&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#attention%EC%9D%98-%EC%9D%98%EC%9D%98&quot; aria-label=&quot;attention의 의의 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Attention의 의의&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;신경망기계번역(NMT)의 성능을 획기적으로 개선하였다.&lt;ul&gt;&lt;li&gt;디코더의 매 time step마다 source의 특정 정보들을 활용할 수 있게 되었기 때문이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;인코더의 마지막 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;만을 디코더에 이용할 수 있으므로 적은 양의 파라미터에 과거의 정보들이 모두 욱여넣어지는 &lt;code&gt;bottlenect problem&lt;/code&gt;을 해결했다.&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 뿐만 아니라 인코더에서 각 time step의 hidden state vector를 모두 활용하기 때문이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Vanishing Gradient 문제를 해결했다.&lt;ul&gt;&lt;li&gt;잘못된 단어를 생성하였을 때 이를 고치기 위해 backpropagation한다고 생각해보자.&lt;ul&gt;&lt;li&gt;먼 거리의 가중치를 업데이트하기 위해서 기존에는 time step을 차례로 거슬러 올라가야 하며, 이 과정에서 정보(gradient)가 소실될 수 있었다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;그러나, Attention이 있을 때는 어떤 time step도 거치지 않고 Attention 모듈의 파라미터를 통해 backpropagation하므로, 먼 거리에 있는 가중치도 정보를 소실하지 않고 바로 업데이트할 수 있게 되었다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;해석가능성을 제공해준다.&lt;ul&gt;&lt;li&gt;Attention distribution 수치를 확인하여 디코더가 인코더의 어느 부분에 집중했는지 확인할 수 있다.&lt;/li&gt;&lt;li&gt;Attention 모듈 신경망은 언제 어떤 단어에 집중해야 하는지도 스스로 학습한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;hr/&gt;&lt;h3 id=&quot;reference&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#reference&quot; aria-label=&quot;reference permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Reference&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://google.github.io/seq2seq/&quot;&gt;Overview - seq2seq&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://www.cs.ox.ac.uk/files/11802/Lecture14NLP6.pdf&quot;&gt;Machine Translation, Seq2seq, &amp;amp; Attention&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[NLP 06 - Beam Search와 BLEU 평가법]]></title><description><![CDATA[Beam Search and BLEU by 주재걸 교수님, BoostCamp AI Tech 4주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/27_beam_search_and_blue/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/27_beam_search_and_blue/</guid><pubDate>Wed, 17 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;beam-search&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#beam-search&quot; aria-label=&quot;beam search permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Beam Search&lt;/h1&gt;&lt;h2 id=&quot;한계-인식&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%95%9C%EA%B3%84-%EC%9D%B8%EC%8B%9D&quot; aria-label=&quot;한계 인식 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;한계 인식&lt;/h2&gt;&lt;h3 id=&quot;seq2seq의-문제점&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#seq2seq%EC%9D%98-%EB%AC%B8%EC%A0%9C%EC%A0%90&quot; aria-label=&quot;seq2seq의 문제점 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Seq2Seq의 문제점&lt;/h3&gt;&lt;p&gt;기존의 seq2seq 모델은 매 time step마다 가장 높은 확률을 가지는 단 하나의 다음 단어를 예측하는 task를 학습한다. 이를 &lt;strong&gt;&lt;code&gt;Greedy Decoding&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;/p&gt;&lt;p&gt;그러나 이 방식은 시퀀스로서의 문장 전체를 보고 예측하는 것이 아니라, 근시안적으로 현재 time step에서 가장 좋아보이는 값을 예측하므로, 잘못 예측한 단어를 한번 내놓으면 되돌릴 수 없다.&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;exhaustive-search&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#exhaustive-search&quot; aria-label=&quot;exhaustive search permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Exhaustive search&lt;/h3&gt;&lt;p&gt;번역의 형태를 수식으로 표현하면 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle mathcolor=&quot;orange&quot;&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;mstyle mathcolor=&quot;purple&quot;&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/munderover&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
P(y|x) &amp;amp;= \textcolor{orange}{P(y_1|x)}\textcolor{purple}{P(y_2|y_1,x)P(y_3|y_2,y_1,x)\dots P(y_T|y_1,\dots,y_{T-1},x)} \\
&amp;amp;= \prod^t_1P(y_t|y_1,\dots,y_{t-1},x)
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:4.8476740000000005em;vertical-align:-2.1738370000000002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.6738370000000002em&quot;&gt;&lt;span style=&quot;top:-5.614398em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.7805610000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.1738370000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.7805610000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.1738370000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.6738370000000002em&quot;&gt;&lt;span style=&quot;top:-5.614398em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.7805610000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:orange&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:orange&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:orange&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:orange&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:orange&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:orange&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:orange&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:purple&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:purple&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:purple&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:purple&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:purple&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:purple&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:purple&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:purple&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:purple&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:purple&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:purple&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot; style=&quot;color:purple&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:purple&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:purple&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em;color:purple&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:purple&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot; style=&quot;color:purple&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:purple&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.328331em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em;color:purple&quot;&gt;T&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:purple&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:purple&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:purple&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:purple&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.1738370000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.7805610000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7805610000000003em&quot;&gt;&lt;span style=&quot;top:-1.882887em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∏&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.267113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.1738370000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;입력 시퀀스 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 주어졌을 때, 가장 그럴듯한 출력 시퀀스 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 얻기 위해서는 &lt;code&gt;joint probability&lt;/code&gt;, 즉 우변의 모든 항들이 다 곱해졌을 때(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msubsup&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\prod^t_1P(y_t|y_1,\dots,y_{t-1},x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.233166em;vertical-align:-0.29971000000000003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∏&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.933456em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2029em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29971000000000003em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;) 가장 큰 값이 나와야 한다.&lt;/p&gt;&lt;p&gt;그러나, 기존의 Seq2Seq 값은 첫번째 단어인 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 확률 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;orange&quot;&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{orange}{P(y_1|x)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:orange&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:orange&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:orange&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:orange&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:orange&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:orange&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:orange&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 구할 때, 뒤의 확률들을 고려할 여력이 없기 때문에, 단순히&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;orange&quot;&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{orange}{P(y_1|x)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:orange&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:orange&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:orange&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:orange&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:orange&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:orange&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:orange&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 최대화시키는 greedy한 방식을 취하게 된다.&lt;/p&gt;&lt;p&gt;그러나 전체적인 관점에서 만약 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;orange&quot;&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{orange}{P(y_1|x)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:orange&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:orange&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:orange&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:orange&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:orange&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:orange&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:orange&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:orange&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 지금 당장은 작더라도, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;purple&quot;&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{purple}{P(y_2|y_1,x)P(y_3|y_2,y_1,x)\dots P(y_T|y_1,\dots,y_{T-1},x)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:purple&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:purple&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:purple&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:purple&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:purple&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:purple&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:purple&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:purple&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:purple&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:purple&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:purple&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot; style=&quot;color:purple&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:purple&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:purple&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em;color:purple&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:purple&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot; style=&quot;color:purple&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:purple&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:purple&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.328331em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em;color:purple&quot;&gt;T&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:purple&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:purple&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:purple&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:purple&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:purple&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 더 커져서 결과적으로 최대의 확률을 얻을 수 있게 만드는 것이 이상적인 방법일 것이다.&lt;/p&gt;&lt;p&gt;이 방법을 어떻게 구현할 수 있을까? 직접 모든 가능한 시퀀스 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 계산해볼까? (&lt;strong&gt;&lt;code&gt;Exhaustive Search&lt;/code&gt;&lt;/strong&gt; - Brute Force 방식)&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이는 매 step &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;마다 vocabulary 크기인 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;V^t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7935559999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7935559999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 가짓수만큼(즉, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;V_n^t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.040556em;vertical-align:-0.247em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7935559999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.22222em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 선택했을때의 확률들) 모두 고려해야한다는 말이다. &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O(V^t)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.043556em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7935559999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 당연히 기하급수적으로 증가하므로 현실적으로 불가능하다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이에 대해 차선책으로 나온 방식이 &lt;strong&gt;&lt;code&gt;Beam Search&lt;/code&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;beam-search-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#beam-search-1&quot; aria-label=&quot;beam search 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Beam Search&lt;/h2&gt;&lt;h3 id=&quot;beam-search란&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#beam-search%EB%9E%80&quot; aria-label=&quot;beam search란 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Beam Search란?&lt;/h3&gt;&lt;p&gt;매 step &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 마다 하나의 단어만을 고려하는 &lt;code&gt;Greedy&lt;/code&gt;와, 매 step &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;마다 모든 단어를 고려하는 &lt;code&gt;Exhaustive&lt;/code&gt;의 절충안으로, &lt;strong&gt;&lt;code&gt;Beam Search&lt;/code&gt;&lt;/strong&gt; 는 정해놓은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{red}{k}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em;color:red&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개만의 단어를 고려한다. 이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개의 경우에 해당하는 decoding output을 &lt;code&gt;hypothesis&lt;/code&gt; 라고 부른다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;code&gt;beam size&lt;/code&gt;라고 부르며, 일반적으로 약 5~10개이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;최대화하고자 하는 beam 값, 즉 hypothesis는 joint probability인데, 이 값에 log를 취하게 되면 각각의 확률들을 덧셈으로 표현할 수 있게 된다. &lt;/p&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msubsup&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msubsup&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;score(y_1,\dots,y_t)= logP_{LM}(y_1,\dots,y_t|x) = \sum^t_{i=1}logP_{LM}(y_i|y_1,\dots,y_{i-1},x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.233166em;vertical-align:-0.29971000000000003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.933456em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2029em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29971000000000003em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;함수는 원래 단조증가하는 형태이기 때문에, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 커지면 커질수록 더 큰 값을 가지게 된다. 따라서, 로그함수를 취한 값으로도 동일하게 hypothesis를 최대화할 수 있다.&lt;/p&gt;&lt;p&gt;이와 같은 방식으로 Beam Search는 globally optimal solution, 전역 최적해를 가져다 주지는 않지만, exhaustive search보다는 훨씬 나은 효율을 보여준다.&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;동작-예시&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%8F%99%EC%9E%91-%EC%98%88%EC%8B%9C&quot; aria-label=&quot;동작 예시 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;동작 예시&lt;/h3&gt;&lt;p&gt;아래의 예시는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 2인 예시이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:730px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:91.40624999999999%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;beam-search1&quot; title=&quot;beam-search1&quot; src=&quot;/static/b607f29e122e2d1d23556ce1280cd0a3/e9beb/beam-search1.png&quot; srcSet=&quot;/static/b607f29e122e2d1d23556ce1280cd0a3/6f3f2/beam-search1.png 256w,/static/b607f29e122e2d1d23556ce1280cd0a3/01e7c/beam-search1.png 512w,/static/b607f29e122e2d1d23556ce1280cd0a3/e9beb/beam-search1.png 730w&quot; sizes=&quot;(max-width: 730px) 100vw, 730px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;여기서 log를 취한 값이 음수가 나옴을 알 수 있는데, log 함수의 입력값이 0과 1사이의 범위를 가지는 확률값이기 때문이다. 음수이더라도, log 함수는 단조증가 형태이므로 더 높은 확률값을 가지면 더 높은 score로 연결됨을 알 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;i.e.  [&amp;#x27;he&amp;#x27;의 확률값 &amp;gt; &amp;#x27;I&amp;#x27;의 확률값] → [-0.7 &amp;gt; -0.9]&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:728px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:104.6875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;beam-search2&quot; title=&quot;beam-search2&quot; src=&quot;/static/9e3ac2bb51a08117a42782b960c9829d/cecac/beam-search2.png&quot; srcSet=&quot;/static/9e3ac2bb51a08117a42782b960c9829d/6f3f2/beam-search2.png 256w,/static/9e3ac2bb51a08117a42782b960c9829d/01e7c/beam-search2.png 512w,/static/9e3ac2bb51a08117a42782b960c9829d/cecac/beam-search2.png 728w&quot; sizes=&quot;(max-width: 728px) 100vw, 728px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;이렇게 매번 score가 가장 높은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개만을 골라 다음 search를 진행하므로, 고려해야할 경우의 수가 너무 크게 늘어나지 않는다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:53.90625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;beam-search3&quot; title=&quot;beam-search3&quot; src=&quot;/static/27410a85c1966abe72cd70fb04a381d3/2bef9/beam-search3.png&quot; srcSet=&quot;/static/27410a85c1966abe72cd70fb04a381d3/6f3f2/beam-search3.png 256w,/static/27410a85c1966abe72cd70fb04a381d3/01e7c/beam-search3.png 512w,/static/27410a85c1966abe72cd70fb04a381d3/2bef9/beam-search3.png 1024w,/static/27410a85c1966abe72cd70fb04a381d3/71c1d/beam-search3.png 1536w,/static/27410a85c1966abe72cd70fb04a381d3/7970d/beam-search3.png 1908w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;이 과정이 마무리 되면, 나온 모든 값들 중 최종 score가 높은 하나의 시나리오를 뽑고, 해당 시나리오를 output으로 하는 decoding을 출력하게 된다.&lt;/p&gt;&lt;h3 id=&quot;stopping-criterion&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#stopping-criterion&quot; aria-label=&quot;stopping criterion permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Stopping Criterion&lt;/h3&gt;&lt;p&gt;Beam Search 디코딩에서, 여러 hyptothesis(시나리오)들은 서로 다른 time step에서 &lt;code&gt;&amp;lt;END&amp;gt;&lt;/code&gt;토큰을 생성하게 된다. 즉, 끝나는 시기가 각각 다르다.&lt;/p&gt;&lt;p&gt;&lt;code&gt;&amp;lt;END&amp;gt;&lt;/code&gt; 토큰을 생성한 hypothesis는 완료 상태가 되어 따로 저장된다. 이후 모든 hypothesis가 끝나면 score를 비교한다.&lt;/p&gt;&lt;h3 id=&quot;finishing-up&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#finishing-up&quot; aria-label=&quot;finishing up permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Finishing up&lt;/h3&gt;&lt;p&gt;최종 score는 결국 다음과 같이 도출된다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/munderover&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/msub&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;score(y_1,\dots,y_t) = \log P_{LM}(y_1,\dots,y_t|x) = \sum^t_{i=1}\log P_LM(y_i|y_1,\dots,y_{i-1},x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.05823em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7805610000000003em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;그런데 score는 joint probability에 log함수를 취하여 음수들의 덧셈연산으로 바꾼 것인데,  확률값은 0과 1 사이이므로 &lt;strong&gt;log 함수값들을 합치면 점점 음수가 더해진다&lt;/strong&gt;. 따라서 길이가 길 수록 더 많은 j, &lt;strong&gt;&lt;div&gt;긴 hyphothesis가 더 낮은 score를 가지는 문제가 생긴다.&lt;/div&gt;&lt;/strong&gt;&lt;/p&gt;&lt;p&gt;따라서 이를 해결하기 위하여 각 hyphotesis를 가지고있는 단어의 개수(즉, 길이)로 나누어 score를 normalize 시켜줄 수 있다.(&lt;code&gt;Normalize by length&lt;/code&gt;)&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mfrac&gt;&lt;/mstyle&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/munderover&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;score(y_1,\dots,y_t) = \textcolor{red}{\frac{1}{t}}\sum^t_{i=1}\log P_{LM}(y_i|y_1,\dots,y_{i-1},x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.05823em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;color:red;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7805610000000003em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;h1 id=&quot;bleu&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#bleu&quot; aria-label=&quot;bleu permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;BLEU&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;BLEU&lt;/code&gt;&lt;/strong&gt;는 자연어 생성모델에서 모델의 품질(결과의 정확도)를 평가하는 방식이다.&lt;/p&gt;&lt;p&gt;기존의 Seq2Seq with attention 등의 모델을 통해 자연어 생성 task를 수행하여 문장을 만들었을 때, target 문장과 비교하여 정확도를 측정할 수 있다.&lt;/p&gt;&lt;p&gt;그런데 만약 time step별로 정답 문장과 예측 문장의 단어들을 비교하면, &lt;strong&gt;중간에 한 단어를 빼먹거나 한 단어를 추가로 잘못 생성하기만 했어도 한칸씩 밀려서 정확도가 굉장히 낮아지게 된다&lt;/strong&gt;. 이는 문장 전체를 비교하고 유사도를 평가하는 것이 아니므로 잘못된 방식이 될 것이다.&lt;/p&gt;&lt;p&gt;따라서 정확도를 측정하기 위해서는, 생성한 sequence 문장과 ground truth 문장 사이의 &lt;strong&gt;전체적인 유사도를 평가할 필요가 있다&lt;/strong&gt;. 이를 위한 방식으로, &lt;strong&gt;&lt;code&gt;precision&lt;/code&gt;&lt;/strong&gt;과 &lt;strong&gt;&lt;code&gt;recall&lt;/code&gt;&lt;/strong&gt; 방식을 생각 해 볼 수 있다.&lt;/p&gt;&lt;h2 id=&quot;precision과-recall&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#precision%EA%B3%BC-recall&quot; aria-label=&quot;precision과 recall permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Precision과 Recall&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;precision(정밀도)&lt;/code&gt;&lt;/strong&gt; : 예측값 중에 얼마나 정답이 많은가? - 예측의 정확도&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;recall(재현율)&lt;/code&gt;&lt;/strong&gt; : 정답의 정보들을 얼마나 빠짐없이 재현했는가? - 정답의 재현도&lt;/p&gt;&lt;p&gt;예를 들어, 다음과 같은 두 문장을 생각해보자.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;Reference&lt;/strong&gt;(정답) : &lt;code&gt;She put my name with yellow hearts&lt;/code&gt;&lt;/p&gt;&lt;p&gt;&lt;strong&gt;Predicted&lt;/strong&gt;(예측) : &lt;code&gt;She took my name and gave yellow letters&lt;/code&gt;&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;#&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;_&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;_&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;_&lt;/mi&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;4&lt;/mn&gt;&lt;mn&gt;8&lt;/mn&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;50&lt;/mn&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;%&lt;/mi&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;#&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;_&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;_&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;_&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;4&lt;/mn&gt;&lt;mn&gt;7&lt;/mn&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;57&lt;/mn&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;%&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;precision = \frac{\#(correct\_words)}{length\_of\_prediction} = \frac{4}{8} = 50\%\\
recall = \frac{\#(correct\_words)}{length\_of\_reference} = \frac{4}{7} = 57\%&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.85396em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.4459999999999997em;vertical-align:-0.996em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.45em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.02778em&quot;&gt;_&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.02778em&quot;&gt;_&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6999999999999997em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;#&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.02778em&quot;&gt;_&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.996em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.00744em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;8&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.80556em;vertical-align:-0.05556em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;%&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.4459999999999997em;vertical-align:-0.996em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.45em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.02778em&quot;&gt;_&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.02778em&quot;&gt;_&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6999999999999997em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;#&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.02778em&quot;&gt;_&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.996em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.00744em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;7&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.80556em;vertical-align:-0.05556em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;7&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;%&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;모델의 품질을 평가하기 위해 precision과 recall이라는 각기 다른 두 값을 평균내서 나타낼수도 있을 것이다.&lt;/p&gt;&lt;p&gt;이 때, 평균의 종류는 3가지가 있는데, 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mtext&gt;산술평균&lt;/mtext&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;/mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mtext&gt;기하평균&lt;/mtext&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mroot&gt;&lt;mrow&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;/mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mroot&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mtext&gt;조화평균&lt;/mtext&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mfrac&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mfrac&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mtext&gt;산술평균&lt;/mtext&gt;&lt;mo&gt;≥&lt;/mo&gt;&lt;mtext&gt;기하평균&lt;/mtext&gt;&lt;mo&gt;≥&lt;/mo&gt;&lt;mtext&gt;조화평균&lt;/mtext&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
산술평균 &amp;amp;= \frac{a+b}{2} \\
기하평균 &amp;amp;= \sqrt[2]{a+b} \\
조화평균 &amp;amp;= \frac{1}{\frac{\frac{1}{a}+\frac{1}{b}}{2}}\\
\end{aligned}\\
산술평균\geq 기하평균 \geq 조화평균&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:6.9551300000000005em;vertical-align:-3.2275650000000002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.7275650000000002em&quot;&gt;&lt;span style=&quot;top:-5.727564999999999em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;산&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;술&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;평&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;균&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.802135em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;기&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;하&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;평&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;균&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.8206949999999997em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;조&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;화&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;평&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;균&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.2275650000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.7275650000000002em&quot;&gt;&lt;span style=&quot;top:-5.727564999999999em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.802135em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord sqrt&quot;&gt;&lt;span class=&quot;root&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.825536em&quot;&gt;&lt;span style=&quot;top:-3.003316em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.93943em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;padding-left:0.833em&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.8994299999999997em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail&quot; style=&quot;min-width:0.853em;height:1.08em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;1.08em&quot; viewBox=&quot;0 0 400000 1080&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M95,702
c-2.7,0,-7.17,-2.7,-13.5,-8c-5.8,-5.3,-9.5,-10,-9.5,-14
c0,-2,0.3,-3.3,1,-4c1.3,-2.7,23.83,-20.7,67.5,-54
c44.2,-33.3,65.8,-50.3,66.5,-51c1.3,-1.3,3,-2,5,-2c4.7,0,8.7,3.3,12,10
s173,378,173,378c0.7,0,35.3,-71,104,-213c68.7,-142,137.5,-285,206.5,-429
c69,-144,104.5,-217.7,106.5,-221
l0 -0
c5.3,-9.3,12,-14,20,-14
H400000v40H845.2724
s-225.272,467,-225.272,467s-235,486,-235,486c-2.7,4.7,-9,7,-19,7
c-6,0,-10,-1,-12,-3s-194,-422,-194,-422s-65,47,-65,47z
M834 80h400000v40h-400000z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.10057000000000005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.8206949999999997em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.11em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.14182em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.14182em&quot;&gt;&lt;span style=&quot;top:-2.6550000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.5508em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen nulldelimiter sizing reset-size3 size6&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8443142857142858em&quot;&gt;&lt;span style=&quot;top:-2.656em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2255000000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line mtight&quot; style=&quot;border-bottom-width:0.049em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.384em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.344em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter sizing reset-size3 size6&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen nulldelimiter sizing reset-size3 size6&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8443142857142858em&quot;&gt;&lt;span style=&quot;top:-2.656em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2255000000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line mtight&quot; style=&quot;border-bottom-width:0.049em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.384em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.344em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter sizing reset-size3 size6&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.345em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.37182em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.14182em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.81882em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.14182em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3768200000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.2275650000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8193em;vertical-align:-0.13597em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;산&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;술&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;평&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;균&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≥&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8193em;vertical-align:-0.13597em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;기&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;하&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;평&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;균&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≥&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;조&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;화&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;평&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;균&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;내분점의 관점에서 보자면, 조화평균과 기하평균은 작은 값에 더 가중치를 많이 두는 방식이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;1:1 내분점(산술평균)보다 작은 쪽에 치우쳐져 있기 때문이다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;f-measure&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#f-measure&quot; aria-label=&quot;f measure permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;F-measure&lt;/h3&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;F&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mfrac&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mn&gt;0.5&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;0.57&lt;/mn&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mn&gt;0.5&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;0.5&lt;/mn&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;0.57&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;0.53&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;F-measure = \frac{precision\times recall}{\frac{1}{2}(precision+recall)} = \frac{0.5\times0.57}{0.5\times(0.5+0.57)} = 0.53&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.76666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;F&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.451548em;vertical-align:-1.080108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3714399999999998em&quot;&gt;&lt;span style=&quot;top:-2.2648919999999997em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.845108em&quot;&gt;&lt;span style=&quot;top:-2.6550000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.394em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.345em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.080108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.25744em;vertical-align:-0.936em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;7&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;7&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;F-measure&lt;/code&gt;&lt;/strong&gt;는 precision과 recall을 이용할 때 조화평균을 이용하는 방식으로, 좀 더 작은 쪽에 가중치를 두어서 평가하는 방식이 된다.&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;그러나 위에서 제시한 품질평가방법들은 &lt;strong&gt;순서를 고려하지 않으므로&lt;/strong&gt;, 정답 단어들이 뒤죽박죽 섞여있더라도 정확하다고 평가하는 오류를 가진다.&lt;/p&gt;&lt;p&gt;이를 해결하기 위해 BLEU 스코어가 나왔다.&lt;/p&gt;&lt;h2 id=&quot;blue-score&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#blue-score&quot; aria-label=&quot;blue score permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;BLUE score&lt;/h2&gt;&lt;h3 id=&quot;blue-score란&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#blue-score%EB%9E%80&quot; aria-label=&quot;blue score란 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;BLUE score란?&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;BiLingual Evaluation Understudy(BLEU)&lt;/code&gt;&lt;/strong&gt;는 기존의 정확도 평가 방식에 추가적으로 순서까지 고려한 평가방식이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;mi&gt;U&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;munder&gt;&lt;mrow&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;_&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;_&lt;/mi&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;_&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;_&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mtext&gt;brevity penalty&lt;/mtext&gt;&lt;/munder&gt;&lt;munder&gt;&lt;munder&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mfrac&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mtext&gt;n-gram&lt;/mtext&gt;&lt;/munder&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;BLEU = \underbrace{min(1,\frac{length\_of\_prediction}{length\_of\_reference})}_{\text{brevity penalty}}\underbrace{(\prod^n_{i=1}precision_i)^{\frac{1}{n}}}_{\text{n-gram}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;U&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:4.2145660000000005em;vertical-align:-2.5631690000000003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3944399999999997em&quot;&gt;&lt;span style=&quot;top:-1.0643319999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.39444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord text mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;brevity penalty&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.3944399999999995em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.39444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.39444em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-1.7504399999999998em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.39444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.39444em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.39444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.39444em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.02778em&quot;&gt;_&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.02778em&quot;&gt;_&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6999999999999997em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.02778em&quot;&gt;_&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.02778em&quot;&gt;_&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.996em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6440000000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.466216em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.651397em&quot;&gt;&lt;span style=&quot;top:-1.2243359999999999em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.6513970000000002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord text mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;n-gram&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6513970000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.6513970000000002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-1.7257280000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.6513970000000002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6513970000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.6513970000000002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∏&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0040200000000001em&quot;&gt;&lt;span style=&quot;top:-3.4130000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen nulldelimiter sizing reset-size3 size6&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8443142857142858em&quot;&gt;&lt;span style=&quot;top:-2.656em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2255000000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line mtight&quot; style=&quot;border-bottom-width:0.049em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.384em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.344em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter sizing reset-size3 size6&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.925669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.5631690000000003em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;&amp;#x27;정답 문장의 단어들과 겹치는 개별 단어가 얼마나 있는가&amp;#x27;를 반영하면서도, 추가적으로 &lt;div&gt;&amp;#x27;n개의 연속된 단어(&lt;strong&gt;&lt;code&gt;n-gram&lt;/code&gt;&lt;/strong&gt;)가 얼마나 정답 문장의 단어연속체와 겹치는가&amp;#x27;를 평가에 반영&lt;/div&gt;한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때 &lt;strong&gt;n은 일반적으로 1-4개 정도&lt;/strong&gt;이다.&lt;/li&gt;&lt;li&gt;precision만 반영하고, recall은 관여되지 않는다.&lt;ul&gt;&lt;li&gt;예시 : The horse is very fast.&lt;/li&gt;&lt;li&gt;위의 문장에서 very가 빠진다고 해서 번역의 질이 크게 떨어지지 않는다.&lt;/li&gt;&lt;li&gt;다만, 어떤 단어(horse)를 아예 다른 단어(morse)로 나타냈을 경우 문장의 의미가 달라져버리는 경우가 생기기 때문에 precision은 고려한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;precision의 곱들을 기하평균 취한다.&lt;ul&gt;&lt;li&gt;산술평균에 비해 작은 쪽에 좀 더 가중치를 주기 위해서이다.&lt;/li&gt;&lt;li&gt;조화평균을 쓰지 않은것은 작은 쪽에 지나치게 큰 가중치를 주기 때문이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;brevity penalty&lt;/code&gt;&lt;/strong&gt;는 &lt;strong&gt;&lt;div&gt;정답 문장에 비해 너무 적은 문장을 생성했을 경우 더 작은 값을 곱해주어 페널티를 주는것&lt;/div&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;예측 문장의 길이가 정답 문장의 길이보다 짧으면, 1보다 작은 수를 곱하므로 정확도가 낮아진다.&lt;/li&gt;&lt;li&gt;반대로 생각해보면, 정답문장의 모든 단어들이 존재하면 예측문장에 존재하면, 예측 문장의 길이가 정답 문장의 길이보다 길더라도, 최대 1밖에 곱해지지 않는다. 그런 점에서 recall의 최댓값으로 1이 나오는것과 동일하게 볼 수 있으므로, recall을 어느정도 고려한 모델이 된다.&lt;/li&gt;&lt;/ul&gt;&lt;hr/&gt;&lt;h3 id=&quot;reference&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#reference&quot; aria-label=&quot;reference permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Reference&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://web.stanford.edu/class/cs224n/slides/cs224n-2019-lecture08-nmt.pdf&quot;&gt;NMT - cs234n&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[NLP 03 - RNN과 이를 활용한 Character Language Modeling]]></title><description><![CDATA[Recurrent Neural Network and Language Modeling by 주재걸 교수님, BoostCamp AI Tech 4주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/24_rnn_character_language_model/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/24_rnn_character_language_model/</guid><pubDate>Tue, 16 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;rnn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#rnn&quot; aria-label=&quot;rnn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;RNN&lt;/h1&gt;&lt;h2 id=&quot;구조&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B5%AC%EC%A1%B0&quot; aria-label=&quot;구조 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;구조&lt;/h2&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:48.04687499999999%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAKCAIAAAA7N+mxAAAACXBIWXMAABYlAAAWJQFJUiTwAAABuElEQVQoz1VSuW7bQBDVN+U78hdpXaYL3KRO6zJAgJQpXKQ14EBWEsmiFJo6Qi4t3hZFxZZILrnknpOlZBv2YDGYnX1v5s1gewCglNIeOfbl5Y/5fGZcj80/5nQy/TkYTA3jxjSHv4dX/f5oOLIs6wg+Wu85EkLUdU3qujkY5xxjXFVV27aEkKIoyrLUeU3mQh5L9FyEgiCwbRsh5Hu+4yBv5YVBeOu6URRVGOui8MKEUKab1w3vyBqRrtfpOo3jKElifQ1Df5NtttlWn02aPjzs4Hk2ACbk4GZbVO0r2XnBKVVKAq04IyIvcwFccCp186Y58rvOUrhhQg6ZR7JOTyYoCLP9fWWNl5Nf1mh8rbsnnmsbxjaONEbKTv+n89O3H9+8//wOF1VHphTiBNBKolX3TPasyMs48+14sVl78nbFkwTu7lSWQUO/G98+fD35cnFGG9o77EBrBoyhLEByUKzbfEMJETUr97BcgP0XFguYzcDzNL7yY8XEq5k1ZzeH3RLuZ/BvCk0Ksu20AmPAOGjNtNWxYDQd9EmRP5FVN7NggCPACZQBlKH2qs5e/oinHkLsHIcfFvYfryMsdLFgbMkAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;rnn_structure&quot; title=&quot;rnn_structure&quot; src=&quot;/static/b70dd3e4924434c596ad794c84b7ca79/2bef9/rnn_structure.png&quot; srcSet=&quot;/static/b70dd3e4924434c596ad794c84b7ca79/6f3f2/rnn_structure.png 256w,/static/b70dd3e4924434c596ad794c84b7ca79/01e7c/rnn_structure.png 512w,/static/b70dd3e4924434c596ad794c84b7ca79/2bef9/rnn_structure.png 1024w,/static/b70dd3e4924434c596ad794c84b7ca79/71c1d/rnn_structure.png 1536w,/static/b70dd3e4924434c596ad794c84b7ca79/da952/rnn_structure.png 1872w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;RNN의 가장 큰 특징은 &lt;strong&gt;매 timestep &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;마다 동일한 function &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 파라미터 셋 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 사용된다&lt;/strong&gt;는 것이다.&lt;/p&gt;&lt;p&gt;매 순간 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;마다 새로운 단어(혹은 형태소)를 입력으로 받아 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t+1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69841em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 연결되는 hidden state &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 내보내는데, 이 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 이용하여 해당 시점에서의 예측값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 내보내는 경우도 있다. 단어의 형태소를 판단하는 RNN의 사례를 들 수 있다.&lt;/p&gt;&lt;p&gt;또, 매 순간 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 출력하지 않고 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 잘 쌓아뒀다가 마지막 순간에 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 단 한번 출력하고 끝나는 경우도 있을것이다. 부정/긍정문을 판단하는 RNN을 예로 들 수 있다.&lt;/p&gt;&lt;p&gt;이 과정에서 사용되는 모든 파라미터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 function &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 매 timestep마다 동일하다. 또한, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 dimension은 hyperparameter로, 모든 timestep에서 동일한 형태를 가진다.&lt;/p&gt;&lt;p&gt;이 때, 이를 수식으로 풀이하면 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;tanh&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t = f_W(h_{t-1},x_t)\\
\begin{aligned}
h_t &amp;amp;= \tanh(\textcolor{blue}{W_{hh}}h_{t-1} + \textcolor{red}{W_{xh}}x_t)\\
y_t &amp;amp;= \textcolor{green}{W_{hy}}h_t
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.0000000000000004em;vertical-align:-1.2500000000000002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7500000000000002em&quot;&gt;&lt;span style=&quot;top:-3.91em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.41em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2500000000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7500000000000002em&quot;&gt;&lt;span style=&quot;top:-3.91em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;tanh&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:blue&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.41em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:green&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2500000000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;여기서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{blue}{W_{hh}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:blue&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{red}{W_{xh}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 어떤 형태인지는 다음 그림을 살펴보도록 하자.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:78.515625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;rnn_transform&quot; title=&quot;rnn_transform&quot; src=&quot;/static/ca3bc1ae93c14410a00638e6325cd16e/2bef9/rnn_transform.png&quot; srcSet=&quot;/static/ca3bc1ae93c14410a00638e6325cd16e/6f3f2/rnn_transform.png 256w,/static/ca3bc1ae93c14410a00638e6325cd16e/01e7c/rnn_transform.png 512w,/static/ca3bc1ae93c14410a00638e6325cd16e/2bef9/rnn_transform.png 1024w,/static/ca3bc1ae93c14410a00638e6325cd16e/f213e/rnn_transform.png 1192w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위의 그림과 같이,  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_{t-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.902771em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 합친 벡터에 적용되는 선형변환 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 두 파트로 나눌 수 있다. 하나는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 적용되어 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 만들 때 사용되는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{red}{W_{xh}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;과 , 나머지 하나는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_{t-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.902771em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 적용되어 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 만들 때 사용되는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{blue}{W_{hh}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:blue&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다. 각각의 파라미터로 선형변환하고 난 뒤, 두 값을 합쳐 비선형변환(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;tanh&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tanh&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;tanh&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)을 적용시켜 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 완성한다.&lt;/p&gt;&lt;p&gt;완성된 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서 추가적인 변환으로 해당 시점의 예측값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 뽑아낼 수도 있다. &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 가는 변환이므로 이 때 사용되는 파라미터를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{green}{W_{hy}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.969438em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:green&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:green&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 한다.&lt;/p&gt;&lt;div&gt;&lt;p&gt;위의 식에서는 편의 상 bias항을 고려하지 않았지만, 실제로 계산할 때에는 bias항이 추가됨을 유의하자.&lt;/p&gt;&lt;/div&gt;&lt;p&gt;여기서 출력된 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 &lt;code&gt;이진 분류(binary classification)&lt;/code&gt;이면 &lt;code&gt;sigmoid&lt;/code&gt;를 적용시키고, &lt;code&gt;다중 분류(multiple classification)&lt;/code&gt;이면 &lt;code&gt;softmax&lt;/code&gt;를 적용시켜 예측값(또는 확률값)으로 나타낼 수 있다.&lt;/p&gt;&lt;p&gt;sigmoid와 softmax의 차이에 대해서는 다음 글을 참조해보자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://m.blog.naver.com/th9231/221989350922&quot;&gt;#수학 : 시그모이드 함수(Sigmoid) / 소프트맥스함수(Softmax)&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;종류&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A2%85%EB%A5%98&quot; aria-label=&quot;종류 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;종류&lt;/h2&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:31.25%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAAA7zAAAO8wEcU5k6AAABIklEQVQY0xWN2XKDIABF/f8/aydtMlk0GAFZVBaFuEUUEkvfzsw9Z24ipCwJ9dtmrC0gIozv+661opw1UkZe15VxFp04hfB2boEIl4QorZLOTqjq9/2Dq+GQSkCe6+a31fXjq2A2BD/P8zgtKbbfN1Grwa3bDZloRk6MfZK6iQ+V1BlEiFXb5jujeVPRRgTvdacqWT8ISyHSxi7O5SW9Iyx0m7S2pRLHmCtyzL8ePPPewyo/wx8sCu/Dg92v+HSnlyP4VkY45zJ6PRUH0dbJYE1L6WffDeetlMM4+uA1LUUOOkp9CKrEEgBNyLPvp3l28yyLogbASpH0WncYx9iWZSfEa1nis8rz5nxuIfTvt7yD5nJRBRynya3uP04z/nu0TfMHZoBITKGOK9UAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;rnn_types&quot; title=&quot;rnn_types&quot; src=&quot;/static/81c22b148c58026c9ee75e031a02f5ce/2bef9/rnn_types.png&quot; srcSet=&quot;/static/81c22b148c58026c9ee75e031a02f5ce/6f3f2/rnn_types.png 256w,/static/81c22b148c58026c9ee75e031a02f5ce/01e7c/rnn_types.png 512w,/static/81c22b148c58026c9ee75e031a02f5ce/2bef9/rnn_types.png 1024w,/static/81c22b148c58026c9ee75e031a02f5ce/9ba38/rnn_types.png 1329w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;One-to-one&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;입력과 출력이 단 하나 → 시퀀스 데이터가 아닌 모델 구조&lt;/li&gt;&lt;li&gt;이전 timestep의 hidden state &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_{t-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.902771em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 들어오지 않고, 입력으로 받은 값을 그대로 출력하므로, 시퀀스 데이터가 아님&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;One-to-many&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;입력이 하나의 timestep으로 이루어지고, 출력이 여러 timestep으로 이루어지는 모델 구조&lt;/li&gt;&lt;li&gt;Task 사례 : &lt;code&gt;Image Captioning&lt;/code&gt;&lt;ul&gt;&lt;li&gt;단일 입력(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 이미지)을 받고, 캡션을 달기 위하여 필요한 단어(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)들을 순차적으로 생성(출력)&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;이 경우 최초의 단일 입력 이후에 빈 입력을 넣기 위하여 같은 사이즈의 zero 벡터/행렬/텐서가 들어가게 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Many-to-one&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;입력이 여러 timestep으로 이루어지고, 출력이 마지막 timestep 하나로 이루어지는 모델 구조&lt;/li&gt;&lt;li&gt;Task 사례 : &lt;code&gt;긍정/부정문 판단&lt;/code&gt;&lt;ul&gt;&lt;li&gt;문장 시퀀스 데이터를 받다가 마지막 순간에 문장의 긍정/부정을 판단하는 이진분류를 수행하여 출력&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Many-to-many&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;입력과 출력이 모두 시퀀스 형태&lt;/li&gt;&lt;li&gt;Task 1 : &lt;code&gt;기계 번역(Machine Translation)&lt;/code&gt;&lt;ul&gt;&lt;li&gt;텍스트(또는 문장)의 입력이 끝나는 시점부터 해석한 정보를 바탕으로 번역된 문장의 단어를 하나씩 출력&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Task 2 : &lt;code&gt;Part-of-Speech(POS) tagging&lt;/code&gt;&lt;ul&gt;&lt;li&gt;각 단어별로 형태소/성분을 분석하므로, 모든 timestep에서 입력과 출력이 다 일어난다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h1 id=&quot;language-model&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#language-model&quot; aria-label=&quot;language model permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Language Model&lt;/h1&gt;&lt;p&gt;언어모델(Language Model)은 주어진 문자열이나 단어의 순서를 바탕으로 다음 단어를 예측하는 task이다.&lt;/p&gt;&lt;p&gt;언어모델링은 철자(character) 단위나 단어(word) 단위 두 가지 경우 모두 수행할 수 있다.&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;character-language-model&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#character-language-model&quot; aria-label=&quot;character language model permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Character Language Model&lt;/h2&gt;&lt;p&gt;철자(character)가 주어졌을 때, 단어내의 다음 철자를 예상하는 언어 모델링을 생각해보자. 이는 매 timestep마다 다음 철자(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)를 출력해야하므로, &lt;code&gt;Many-to-many&lt;/code&gt; task에 해당한다.&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;과정&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B3%BC%EC%A0%95&quot; aria-label=&quot;과정 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;과정&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:902px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:80.46875000000001%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;character_language_modeling&quot; title=&quot;character_language_modeling&quot; src=&quot;/static/a94bf62d8f6935855a923fc3f8980fe9/58213/character_language_modeling.png&quot; srcSet=&quot;/static/a94bf62d8f6935855a923fc3f8980fe9/6f3f2/character_language_modeling.png 256w,/static/a94bf62d8f6935855a923fc3f8980fe9/01e7c/character_language_modeling.png 512w,/static/a94bf62d8f6935855a923fc3f8980fe9/58213/character_language_modeling.png 902w&quot; sizes=&quot;(max-width: 902px) 100vw, 902px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ol&gt;&lt;li&gt;학습 시퀀스 데이터에서 unique한 철자를 모아 사전을 구축한다.&lt;/li&gt;&lt;li&gt;해당 철자들을 &lt;code&gt;len(vocab)&lt;/code&gt;만큼의 차원을 가지는 원-핫 벡터로 변환한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 만들어내기 위해 변환을 수행한다. 이 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_0&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 hidden state의 dimension을 가지는 영벡터/행렬/텐서로 한다.&lt;/li&gt;&lt;li&gt;선형변환(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_{hy}h_t+b&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.980548em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;) 시킨 후, 다중 분류이므로 Softmax를 적용하여 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 내보낸다. &lt;ul&gt;&lt;li&gt;이 때 선형변환된 값을 Logit으로 여긴다. output layer에서 나온 값 중 하나를 골라 다음에 나올 철자로 예측하기 때문이다.&lt;ul&gt;&lt;li&gt;Logit : log와 odds의 합성으로, 딥러닝에서는 일반적인 0~1 구간의 확률값이 아니라 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∞&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;-\infin&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.66666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∞&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; ~ &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∞&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\infin&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∞&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 의 범위에 있는 날확률값을 의미한다. 다중 분류에서 Softmax 함수의 입력값으로 자주 사용된다.&lt;/li&gt;&lt;li&gt;Logit의 개념을 잘 모르겠다면 &lt;a href=&quot;https://haje01.github.io/2019/11/19/logit.html#:~:text=%EC%A6%89%2C%20%EC%98%A4%EC%A6%88%EC%97%90%20%EC%9E%90%EC%97%B0%EB%A1%9C%EA%B7%B8,log%20%2B%20odds%EC%97%90%EC%84%9C%20%EB%82%98%EC%98%A8%20%EB%A7%90.&amp;amp;text=%ED%99%95%EB%A5%A0%EC%9D%98%20%EB%B2%94%EC%9C%84%EB%8A%94%20%5B0,%E2%88%92%E2%88%9E%2C%E2%88%9E%5D%20%EC%9D%B4%EB%8B%A4.&quot;&gt;로짓(Logit)이란?&lt;/a&gt;을 참고하자.&lt;/li&gt;&lt;li&gt;&lt;a href=&quot;https://opentutorials.org/module/3653/22995#:~:text=logit%EC%9D%80%20&amp;#x27;logistic&amp;#x27;%20%EA%B3%BC%20%2B,%EA%B7%B8%EB%8C%80%EB%A1%9C%20%EA%B0%80%EC%A0%B8%EC%98%A8%EB%8B%A4%EA%B3%A0%20%EB%B3%B4%EC%8B%9C%EB%A9%B4%20%EB%90%98%EA%B2%A0%EC%8A%B5%EB%8B%88%EB%8B%A4.&quot;&gt;logit과 sigmoid, softmax의 관계&lt;/a&gt;도 참고하자.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;첫 출력을 보면 &amp;#x27;h&amp;#x27; 다음 &amp;#x27;e&amp;#x27;가 와야하는데 실제로는 &amp;#x27;o&amp;#x27;가 4.1을 받아 출력되었음을 알 수 있는데, 이와 같은 잘못된 결과에 페널티를 주고 올바른 결과인 &amp;#x27;e&amp;#x27;가 더 높은 값을 얻도록 하기 위해 Softmax-loss를 적용하여 업데이트한다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;해당 모델의 학습결과는 여기서 확인하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;http://karpathy.github.io/2015/05/21/rnn-effectiveness/&quot;&gt;The Unreasonable Effectiveness of Recurrent Neural Networks&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;rnn의-역전파-알고리즘-bptt&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#rnn%EC%9D%98-%EC%97%AD%EC%A0%84%ED%8C%8C-%EC%95%8C%EA%B3%A0%EB%A6%AC%EC%A6%98-bptt&quot; aria-label=&quot;rnn의 역전파 알고리즘 bptt permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;RNN의 역전파 알고리즘, BPTT&lt;/h3&gt;&lt;p&gt;RNN에서의 순전파는 각 timestep마다 Loss를 모두 계산하고, 역전파(Backpropagation through time, BPTT)는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(W_{hh},W_{xh},W_{hy})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 모두 연산해야하므로, 단어/철자 수가 수천/수만을 넘어가는 텍스트 데이터에서는 계산할 양이 매우 방대해진다.&lt;/p&gt;&lt;p&gt;따라서 이를 현실적으로 수행하기 어렵기 때문에, 시퀀스 데이터의 일부를 잘라내어 한번에 학습할 수 있는 길이를 제한시키는 &lt;strong&gt;&lt;code&gt;truncation&lt;/code&gt;&lt;/strong&gt;을 사용한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:48.828125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;truncated_bptt&quot; title=&quot;truncated_bptt&quot; src=&quot;/static/ec973c15488db36065c203392260ffad/2bef9/truncated_bptt.png&quot; srcSet=&quot;/static/ec973c15488db36065c203392260ffad/6f3f2/truncated_bptt.png 256w,/static/ec973c15488db36065c203392260ffad/01e7c/truncated_bptt.png 512w,/static/ec973c15488db36065c203392260ffad/2bef9/truncated_bptt.png 1024w,/static/ec973c15488db36065c203392260ffad/71c1d/truncated_bptt.png 1536w,/static/ec973c15488db36065c203392260ffad/6fa81/truncated_bptt.png 1856w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;searching-for-interpretable-cells&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#searching-for-interpretable-cells&quot; aria-label=&quot;searching for interpretable cells permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Searching for Interpretable Cells&lt;/h3&gt;&lt;p&gt;그렇다면, RNN 모델에서 시퀀스 데이터에서 문맥을 이해하고, 해석가능한(interpretable) 정보는 어느 셀에 담겨있을까?&lt;/p&gt;&lt;p&gt;RNN에서 필요로하는 정보를 저장하는 공간은 결국 매 timestep마다 업데이트되는 hidden state vector,  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:26.171875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAFCAIAAADKYVtkAAAACXBIWXMAABYlAAAWJQFJUiTwAAABCklEQVQY022N20vCUADGz/8bJfTS08IeDGyw5eNuOQl7KRjWLl6ybSKaFsswV7tRcbB3s7Qdt51OBb3Ux48fH3wPHzCcYFu5Kmg3eXVEKUOm2mIPG6WjNl3WaEndE2q0dMYcqKysMbLOysZ+pV6qGMfKhTdwgC1XT7Z29HxRpXZ1qtDLbV5u5Ppr66Mi7Zfle16c8KIriC4vjTmecMcJhFtOmAgSmDYbUOQf662Oblq1pnnatnXL0qzOeW/QdfpdxzaH7vV4EYVvYTAP/S8HxOEs8AFCKEvTBOMlxh8Yx99+T7M5SmbL+DVekbJYkf2fgCT5GbJfIxSHDx58foo8L/L9FwinEJKDv3wCYEryFK6Z7JYAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;quote_dection_cell&quot; title=&quot;quote_dection_cell&quot; src=&quot;/static/e74e14570a3f3d9e0c46038ae4d4a936/2bef9/quote_detection_cell.png&quot; srcSet=&quot;/static/e74e14570a3f3d9e0c46038ae4d4a936/6f3f2/quote_detection_cell.png 256w,/static/e74e14570a3f3d9e0c46038ae4d4a936/01e7c/quote_detection_cell.png 512w,/static/e74e14570a3f3d9e0c46038ae4d4a936/2bef9/quote_detection_cell.png 1024w,/static/e74e14570a3f3d9e0c46038ae4d4a936/71c1d/quote_detection_cell.png 1536w,/static/e74e14570a3f3d9e0c46038ae4d4a936/3126c/quote_detection_cell.png 1876w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위의 이미지는 학습시킨 RNN 모델에서 특정 hidden state vector의 dimension을 고정해놓고, 해당 dimension 값의 변화를 추적한 이미지이다. 따옴표(quote)를 기준으로 해당 dimension의 hidden state 값이 확연히 바뀌는 것을 알 수 있는데, 이는 hidden state가 따옴표가 열렸는지 닫혔는지를 해석하는 방식을 학습했다고 이해할 수 있다.&lt;/p&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;rnn-모델의-문제점&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#rnn-%EB%AA%A8%EB%8D%B8%EC%9D%98-%EB%AC%B8%EC%A0%9C%EC%A0%90&quot; aria-label=&quot;rnn 모델의 문제점 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;RNN 모델의 문제점&lt;/h2&gt;&lt;h3 id=&quot;vanishingexploding-gradient-problem&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#vanishingexploding-gradient-problem&quot; aria-label=&quot;vanishingexploding gradient problem permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Vanishing/Exploding Gradient Problem&lt;/h3&gt;&lt;p&gt;앞서 살펴본 (vanilla, original) RNN 모델은 훌륭하게 문맥을 파악했다. 그러나, 해당 RNN 모델의 학습 방식에는 한계가 있다.&lt;/p&gt;&lt;p&gt;매 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;마다 동일한 matrix &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 곱해주기때문에, 역전파 과정에서 &lt;strong&gt;&lt;code&gt;기울기 소실(Vanishing Gradient)&lt;/code&gt;&lt;/strong&gt;이나 &lt;strong&gt;&lt;code&gt;기울기 폭발(Exploding Gradient)&lt;/code&gt;&lt;/strong&gt;이 생기게 된다. 이 경우 gradient가 뒤로 전파될수록 영향력이 사라지거나, 너무 커다란 영향을 미치게 된다. 따라서, &lt;code&gt;long-term dependancy&lt;/code&gt; 문제가 제시된다.&lt;/p&gt;&lt;p&gt;이 문제를 해결하기 위해 &lt;code&gt;LSTM&lt;/code&gt;과 &lt;code&gt;GRU&lt;/code&gt;같은 RNN의 발전된 형태가 나오게 되었다.&lt;/p&gt;&lt;hr/&gt;&lt;h3 id=&quot;reference&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#reference&quot; aria-label=&quot;reference permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Reference&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;http://cs231n.stanford.edu/slides/2017/cs231n_2017_lecture10.pdf&quot;&gt;cs231n_2017_lecture10.pdf&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;http://karpathy.github.io/2015/05/21/rnn-effectiveness/&quot;&gt;The Unreasonable Effectiveness of Recurrent Neural Networks&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[NLP 04 - RNN을 개선한 LSTM과 GRU 모델]]></title><description><![CDATA[LSTM and GRU by 주재걸 교수님, BoostCamp AI Tech 4주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/25_lstm_gru copy/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/25_lstm_gru copy/</guid><pubDate>Tue, 16 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;lstm--gru&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#lstm--gru&quot; aria-label=&quot;lstm  gru permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;LSTM &amp;amp; GRU&lt;/h1&gt;&lt;h2 id=&quot;lstm&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#lstm&quot; aria-label=&quot;lstm permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;LSTM&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Long Short-Term Memory(LSTM)&lt;/code&gt;&lt;/strong&gt;은 기존 RNN모델에서 &lt;strong&gt;&lt;code&gt;Vanishing/Exploding Gradient&lt;/code&gt;&lt;/strong&gt; 문제를 해결하고, &lt;strong&gt;&lt;code&gt;long-term depandancy&lt;/code&gt;&lt;/strong&gt; 문제를 개선한 모델이다.&lt;/p&gt;&lt;p&gt;hidden state를 단기(short-term)기억소자로 볼 수 있고, 보다 먼 timestep(long-term)의 정보까지 잘 반영하도록 만들었기 때문에 이러한 이름이 붙었다.&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;구조&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B5%AC%EC%A1%B0&quot; aria-label=&quot;구조 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;구조&lt;/h3&gt;&lt;p&gt;기존의 RNN 모델이 다음과 같은 형태였다면,&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t = f_w(x_t,h_{t-1})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;LSTM 모델의 형태는 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;{&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;}&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\{C_t,h_t\} = LSTM(x_t,C_{t-1},h_{t-1})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : &lt;strong&gt;Cell State Vector&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : &lt;strong&gt;Hidden State Vector&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:37.5%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;lstm-chain&quot; title=&quot;lstm-chain&quot; src=&quot;/static/2795bc16b012322f7767cd4d940ba2e3/2bef9/lstm-chain.png&quot; srcSet=&quot;/static/2795bc16b012322f7767cd4d940ba2e3/6f3f2/lstm-chain.png 256w,/static/2795bc16b012322f7767cd4d940ba2e3/01e7c/lstm-chain.png 512w,/static/2795bc16b012322f7767cd4d940ba2e3/2bef9/lstm-chain.png 1024w,/static/2795bc16b012322f7767cd4d940ba2e3/71c1d/lstm-chain.png 1536w,/static/2795bc16b012322f7767cd4d940ba2e3/a878e/lstm-chain.png 2048w,/static/2795bc16b012322f7767cd4d940ba2e3/4942b/lstm-chain.png 2233w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;이 중 &lt;strong&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 좀 더 완전하고 통합적인 정보를 제공&lt;/strong&gt;한다. &lt;strong&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 한번 더 가공(필터링)해서 해당 timestep에서 노출할 필요가 있는 정보들만 남긴 형태&lt;/strong&gt;라고 볼 수 있다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:29.296875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAABE0lEQVQY03WQ22rDMBBE/f8/VUgbmpASKMWJ1Vix5TuxLFl3I1tS1ea5wz4sC7NndhPGGJ3nCaFNmxDCYsw0YUKIEGKxthuJ9z78oySahZTX80HQyXmvlKrrGkL4eIxaqwKB4H3dO1i7svOw2YrOV0OATSBzSCgleMR5k6tFxWVa66qqsizr+mHRHNef3vndme6v7JTr47d8z/gH1PuLuVXrL5lzbox5JolN27YRjqdp0XKAqXPu5YTfLvMByNeUxjoXyy7VoLSRTOOFEfg0SymHPzHOtFRleovDZrA5MjnSX4AUrUWtvdeWy5BAeAcAEEKf5q7vMcarteu6SiFQibZtiz/jjCrJZzqNj14r4TYbgv8B7oBOn8+WicUAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;lstm_gates&quot; title=&quot;lstm_gates&quot; src=&quot;/static/cc43e293065452a61da12fc9751d7438/2bef9/lstm-gates.png&quot; srcSet=&quot;/static/cc43e293065452a61da12fc9751d7438/6f3f2/lstm-gates.png 256w,/static/cc43e293065452a61da12fc9751d7438/01e7c/lstm-gates.png 512w,/static/cc43e293065452a61da12fc9751d7438/2bef9/lstm-gates.png 1024w,/static/cc43e293065452a61da12fc9751d7438/71c1d/lstm-gates.png 1536w,/static/cc43e293065452a61da12fc9751d7438/10d53/lstm-gates.png 1724w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;LSTM은 &lt;strong&gt;주어진 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_{t-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.902771em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 바탕으로 4개의 게이트(gate)를 이용해 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_{t-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.891661em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 수정하고 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 계산&lt;/strong&gt;한다.&lt;/p&gt;&lt;p&gt;먼저, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_{t-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.902771em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 붙여 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 이용하여 &lt;strong&gt;벡터로 선형변환&lt;/strong&gt;시키고, 변환시킨 벡터를 4개로 쪼개어 각각 sigmoid와 tanh를 취해 &lt;strong&gt;4개의 게이트&lt;/strong&gt;를 만든다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;i&lt;/code&gt;&lt;/strong&gt; : &lt;code&gt;Input gate&lt;/code&gt;, 셀에 해당 정보를 써 넣을 것인가?&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i_t = \sigma(W_i\cdot[h_{t-1},x_t] + b_i)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.80952em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;f&lt;/code&gt;&lt;/strong&gt; : &lt;code&gt;Forget gate&lt;/code&gt;, 셀에서 정보를 얼마나 잊을 것인가? 거꾸로 말하자면, 얼마나 보존할 것인가?&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f_t = \sigma(W_f\cdot[h_{t-1},x_t] + b_f)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;o&lt;/code&gt;&lt;/strong&gt; : &lt;code&gt;Output gate&lt;/code&gt;, 현재 Cell(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)를 출력값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 가공할 때 얼마나 필터링 할 것인가?&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;o_t = \sigma(W_O[h_{t-1},x_t]+b_O)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;o_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 sigmoid를 거치므로  0~1 사이의 값이 된다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;⊙&lt;/mo&gt;&lt;mi&gt;tanh&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t =o_t \odot \tanh(C_t)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.73333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⊙&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;tanh&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;최종 출력값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 tanh를 거친 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 0과 1사이의 적절한 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;o_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 element-wise product함으로써  Cell state 정보를 일정 비율로 작게 만들어(filtering) 내보내는 값이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;div&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 이전까지의 모든 과거 정보를 담은 벡터, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 해당 time step에서 예측값에 사용하기 위해 지금 당장 필요한 정보만을 (&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로부터) 필터링한 값이다.&lt;/div&gt;&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;g&lt;/code&gt;&lt;/strong&gt; : &lt;code&gt;Gate gate&lt;/code&gt;, 현재 cell에 해당 정보를 얼마나(how much) 기록할것인가?&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;tanh&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;⊙&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;⊙&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
\tilde{C_t}&amp;amp;=\tanh(W_C\cdot[h_{t-1},x_t]+b_c)\\
C_t &amp;amp;= f_t \odot C_{t-1} + i_t \odot \tilde{C_t}
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.16038em;vertical-align:-1.33019em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.83019em&quot;&gt;&lt;span style=&quot;top:-3.91em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9201899999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6023300000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.32981em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.33019em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.83019em&quot;&gt;&lt;span style=&quot;top:-3.91em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;tanh&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.32981em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⊙&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⊙&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9201899999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6023300000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.33019em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tilde C_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0701899999999998em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9201899999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6023300000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.16666em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 Gate gate 값이며, 현재 timestep에서 만들어진 정보를 의미한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;⊙&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i_t \odot \tilde{C_t}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.80952em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⊙&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0701899999999998em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9201899999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6023300000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 현재 정보를 기억할 게이트 값으로, Gate gate 값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tilde{C_t}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0701899999999998em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9201899999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6023300000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 Input gate 값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.80952em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;과 elemen-wise product된다.&lt;ul&gt;&lt;li&gt;이 때, Gate gate 값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tilde{C_t}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0701899999999998em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9201899999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6023300000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;은 tanh연산을 거쳤으므로 범위가 -1~1이다. 따라서, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;⊙&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i_t \odot \tilde{C_t}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.80952em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⊙&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0701899999999998em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9201899999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6023300000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 Cell state에 반영시키되, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tilde{C_t}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0701899999999998em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9201899999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6023300000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 반영의 방향(증감)을 정한다고 볼 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;이후, 이전 Cell state가 forget gate 값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 더해져 새로운 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 만든다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;sigmoid 사용 : &lt;strong&gt;&lt;code&gt;i&lt;/code&gt;, &lt;code&gt;f&lt;/code&gt;, &lt;code&gt;o&lt;/code&gt;&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;0과 1 사이의 값을 가지는 sigmoid의 특성 상 기존의 벡터 값을 일정 부분만 떼내어 보존하는 역할을 한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;tanh 사용 : &lt;strong&gt;&lt;code&gt;g&lt;/code&gt;&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;-1과 1 사이의 값을 가지는 tanh의 특성 상 현재 timestep에서 계산되는 유의미한 정보(증/감)를 담는 역할을 한다.(vanilla rnn에서의 마지막 tanh와 같음)&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;gru&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gru&quot; aria-label=&quot;gru permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GRU&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Gated Recurrent Unit(GRU)&lt;/code&gt;&lt;/strong&gt;은 LSTM의 모델 구조를 경량화하여 메모리 요구량을 줄이고 빠른 학습이 가능하도록 만든 모델이다.&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;구조-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B5%AC%EC%A1%B0-1&quot; aria-label=&quot;구조 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;구조&lt;/h3&gt;&lt;p&gt;기존의 LSTM이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 이원화하여 가져가는 모델이었다면, &lt;strong&gt;GRU는 두 state vector를 합쳐 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 하나만 남겨놓았다&lt;/strong&gt;. 이때, 일원화된 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 LSTM에서의 핵심적인 역할을 하던 Cell state vector &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 의미적으로 더 유사한 값이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:30.859375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;gru&quot; title=&quot;gru&quot; src=&quot;/static/f2716bc289734d8b545926b38a224692/2bef9/gru.png&quot; srcSet=&quot;/static/f2716bc289734d8b545926b38a224692/6f3f2/gru.png 256w,/static/f2716bc289734d8b545926b38a224692/01e7c/gru.png 512w,/static/f2716bc289734d8b545926b38a224692/2bef9/gru.png 1024w,/static/f2716bc289734d8b545926b38a224692/71c1d/gru.png 1536w,/static/f2716bc289734d8b545926b38a224692/332b4/gru.png 1826w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;msub&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;/mstyle&gt;&lt;mspace width=&quot;1em&quot;&gt;&lt;/mspace&gt;&lt;mtext&gt;in LSTM&lt;/mtext&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;/mstyle&gt;&lt;mspace width=&quot;1em&quot;&gt;&lt;/mspace&gt;&lt;mtext&gt;in GRU&lt;/mtext&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
C_t &amp;amp;= \textcolor{red}{f_t}\cdot C_{t-1}+\textcolor{blue}{i_t}\cdot \textcolor{green}{\tilde{C_t}} \quad \textnormal{in LSTM} \\
h_t &amp;amp;= \textcolor{red}{(1-z_t)}\cdot h_{t-1} + \textcolor{blue}{z_t} \cdot \textcolor{green}{\tilde{h_t}} \quad \textnormal{in GRU}
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.17149em;vertical-align:-1.335745em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.835745em&quot;&gt;&lt;span style=&quot;top:-3.9155550000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.324255em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.335745em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.835745em&quot;&gt;&lt;span style=&quot;top:-3.9155550000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em;color:red&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9201899999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em;color:green&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6023300000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:1em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;in LSTM&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.324255em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot; style=&quot;color:red&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em;color:red&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em;color:blue&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9312999999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.61344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:1em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;in GRU&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.335745em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;기존 LSTM에서는 input gate와 forget gate 값을 이용했다면, GRU에서는 input gate에 해당하는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;만 남겨두고, &lt;strong&gt;&lt;div&gt;forget gate &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(1-z_t)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 바꾸어 처리&lt;/div&gt;&lt;/strong&gt;한다. &lt;/p&gt;&lt;p&gt;현재 값인 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tilde{h_t}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0812999999999997em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9312999999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.61344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와  과거의 정보를 통합한 값인 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_{t-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.902771em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 비율을 조정하여, 합쳐서 1이 되도록 만들면서, &lt;strong&gt;현재와 과거 값 간의 가중평균&lt;/strong&gt;을 내게 된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;div&gt;이 과정에서 게이트도 하나(fortget gate) 줄이게 되어, 계산양과 메모리 요구량을 줄여 경량화&lt;/div&gt;&lt;/strong&gt;하였다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;lstm과-gru의-역전파&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#lstm%EA%B3%BC-gru%EC%9D%98-%EC%97%AD%EC%A0%84%ED%8C%8C&quot; aria-label=&quot;lstm과 gru의 역전파 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;LSTM과 GRU의 역전파&lt;/h2&gt;&lt;p&gt;기존의 RNN은 역전파 과정에서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_{hh}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 계속 곱해주었기 때문에 Vanishing/Exploding Gradient 문제가 생겼었다.&lt;/p&gt;&lt;p&gt;그러나,  LSTM(과 GRU)의 역전파는 &lt;strong&gt;Cell State Vector &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_{t-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.891661em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 그 때 그 때 다른 값으로 이루어지는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.10764em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 곱하고, 필요로 하는 정보 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;~&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tilde{C_t}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0701899999999998em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9201899999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.6023300000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;~&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 관련 항을 더해주어 Vanishing/Exploding Gradient를 막았다.&lt;/strong&gt;&lt;/p&gt;&lt;p&gt;또한, 이러한 &lt;strong&gt;덧셈 연산이 역전파과정에서 Gradient를 복사하여 매번 전달해주는 역할&lt;/strong&gt;을 하므로,, RNN에 비해 멀리있는 timestep의 gradient까지 큰 변형없이 전달하게 되어 결과적으로 &lt;strong&gt;long-term depandency를 어느정도 확보&lt;/strong&gt;하였다.
다만, LSTM과 GRU는 Long Term Dependency 문제를 완벽하게 해결하지는 못했다. 멀리 있는 정보는 곱해져 최근으로 오면서 점점 소실되게 된다.&lt;/p&gt;&lt;p&gt;2017년 등장한 attention 구조를 활용한 transformer가 이러한 문제를 근본적으로 해결하게 되었다.&lt;/p&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;rnnlstmgru-요약&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#rnnlstmgru-%EC%9A%94%EC%95%BD&quot; aria-label=&quot;rnnlstmgru 요약 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;RNN/LSTM/GRU 요약&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;RNN&lt;/code&gt;은 다양한 길이를 가질 수 있는 &lt;strong&gt;시퀀스 데이터에 특화된 딥러닝 모델&lt;/strong&gt;이다.&lt;ul&gt;&lt;li&gt;아키텍쳐 디자인에 유연성을 제공한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;Vanilla RNN&lt;/code&gt;은 구조가 간단하지만 Vanishing/Exploding Gradient 문제 때문에 실제로 많이 사용되지는 않는다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;LSTM&lt;/code&gt;과 &lt;code&gt;GRU&lt;/code&gt;는 실제 많이 사용하고 있으며, Vanishing/Exploding Gradient 문제를 &lt;strong&gt;덧셈연산으로 해결한 모델&lt;/strong&gt;이다.&lt;ul&gt;&lt;li&gt;&lt;code&gt;Cliping Gradient&lt;/code&gt; 같은 방법으로 해결할 수도 있다고 한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;hr/&gt;&lt;h3 id=&quot;reference&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#reference&quot; aria-label=&quot;reference permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Reference&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;http://colah.github.io/posts/2015-08-Understanding-LSTMs/&quot;&gt;Understanding LSTM Networks&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://imgur.com/gallery/vaNahKE&quot;&gt;RNN vs LSTM: Vanishing Gradients&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[NLP 01 - NLP와 Bag-of-Words란?]]></title><description><![CDATA[Intro to LNP,Bag-of-Words by 주재걸 교수님, BoostCamp AI Tech 4주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/22_bag-of-words/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/22_bag-of-words/</guid><pubDate>Mon, 15 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;intro-to-nlp&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#intro-to-nlp&quot; aria-label=&quot;intro to nlp permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Intro to NLP&lt;/h1&gt;&lt;h2 id=&quot;academic-disciplines-related-to-nlp&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#academic-disciplines-related-to-nlp&quot; aria-label=&quot;academic disciplines related to nlp permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Academic Disciplines related to NLP&lt;/h2&gt;&lt;h3 id=&quot;nlp&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#nlp&quot; aria-label=&quot;nlp permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;NLP&lt;/h3&gt;&lt;p&gt;major conferences :  &lt;code&gt;ACL&lt;/code&gt;, &lt;code&gt;EMNLP&lt;/code&gt;, &lt;code&gt;NAACL&lt;/code&gt;&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;NLP&lt;/code&gt;&lt;/strong&gt;란? - &lt;strong&gt;&lt;code&gt;Natural Language Processing, 자연어처리&lt;/code&gt;&lt;/strong&gt;&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;자연어처리가-다루는-task&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9E%90%EC%97%B0%EC%96%B4%EC%B2%98%EB%A6%AC%EA%B0%80-%EB%8B%A4%EB%A3%A8%EB%8A%94-task&quot; aria-label=&quot;자연어처리가 다루는 task permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;자연어처리가 다루는 Task&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;Low-level parsing&lt;/code&gt;&lt;ul&gt;&lt;li&gt;Tokenization(문장을 단어단위로 쪼개는 것), stemming(단어의 어근 추출)&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Word and phrase level&lt;ul&gt;&lt;li&gt;&lt;code&gt;Named entity recognition(NER)&lt;/code&gt; - 여러 단어의 조합을 하나의 고유명사로 인식 i.e. New York Times&lt;/li&gt;&lt;li&gt;&lt;code&gt;part-of-speech(POS) tagging&lt;/code&gt; - 문장 내에서 단어의 품사나 성분 파악&lt;/li&gt;&lt;li&gt;&lt;code&gt;noun-phrase chunking&lt;/code&gt;, &lt;code&gt;dependency parsing&lt;/code&gt;, &lt;code&gt;coreference resolution&lt;/code&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Sentence level&lt;ul&gt;&lt;li&gt;&lt;code&gt;Sentiment analysis&lt;/code&gt; - 문장의 긍정/부정여부, 감정 등을 파악&lt;/li&gt;&lt;li&gt;&lt;code&gt;machine translation&lt;/code&gt; - 타 언어로 단어의 어감과 어순을 살려 번역&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Multi-sentence and paragraph level&lt;ul&gt;&lt;li&gt;&lt;code&gt;Entailment prediction&lt;/code&gt; - 두 문장 간의 내포관계, 모순관계 등을 예측&lt;/li&gt;&lt;li&gt;&lt;code&gt;question answering&lt;/code&gt; - 질문의 대답에 해당하는 문장을 선택 i.e. 구글 검색시 미리보기&lt;/li&gt;&lt;li&gt;&lt;code&gt;dialog systems&lt;/code&gt; - 챗봇처럼 대화를 수행&lt;/li&gt;&lt;li&gt;&lt;code&gt;summarization&lt;/code&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;text-mining&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#text-mining&quot; aria-label=&quot;text mining permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Text Mining&lt;/h3&gt;&lt;p&gt;major conference : &lt;code&gt;KDD&lt;/code&gt;, &lt;code&gt;The WebConf(WWW)&lt;/code&gt;, &lt;code&gt;WSDM&lt;/code&gt;, &lt;code&gt;CIKM&lt;/code&gt;, &lt;code&gt;ICWSM&lt;/code&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;텍스트와 문서 데이터에서 유용한 정보 추출 - i.e. 방대한 뉴스데이터에서 특정 키워드에 연관된 트렌드 분석&lt;/li&gt;&lt;li&gt;&lt;code&gt;Document clustering(topic modeling, 문서 군집화)&lt;/code&gt; - i.e. 뉴스데이터를 모아 주제별로 군집화&lt;/li&gt;&lt;li&gt;Computational social science - 빅데이터 기반 사회과학에 연관 i.e. 트위터 키워드 분석한 사회상 예측&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;information-retrieval-정보-검색&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#information-retrieval-%EC%A0%95%EB%B3%B4-%EA%B2%80%EC%83%89&quot; aria-label=&quot;information retrieval 정보 검색 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Information retrieval 정보 검색&lt;/h3&gt;&lt;p&gt;major conferences: &lt;code&gt;SIGAR&lt;/code&gt;, &lt;code&gt;WSDM&lt;/code&gt;, &lt;code&gt;CIKM&lt;/code&gt;, &lt;code&gt;RecSys&lt;/code&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;빅데이터 기반 사회과학에 연관이 있다.&lt;/li&gt;&lt;li&gt;자연어처리나 텍스트마이닝에 비하여 상대적으로 성숙하였기 때문에 연구가 비교적 활발하지 않다.&lt;/li&gt;&lt;li&gt;최근 핫한 분야인 추천시스템 쪽으로 발달하고 있다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;nlp-트렌드&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#nlp-%ED%8A%B8%EB%A0%8C%EB%93%9C&quot; aria-label=&quot;nlp 트렌드 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;NLP 트렌드&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Word Embedding&lt;/code&gt;&lt;/strong&gt; - 주어진 텍스트 데이터를 단어단위로 분리하고, 각 단어를 다차원으로 이루어진 vector로 표현하는 것.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;RNN-family models(LSTM and GRUs)&lt;/code&gt;&lt;/strong&gt; - 시퀀스 데이터인 문장을 처리하는데에 특화된 RNN을 사용하는 모델. 자연어처리의 핵심 모델&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Attention 모듈과 Transformer 모델&lt;/code&gt;&lt;/strong&gt; - 기존의 RNN을 self-attention으로 완전히 대체하며 큰 성능 향상을 이루었음. 2017년에 발표.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Transformer 모델처럼, 대부분의 NLP모델은 기계번역(machine translation)을 수행하기 위해 발달된 형태이다.&lt;/p&gt;&lt;p&gt;과거에는 특정 NLP task를 처리하기 위한 특화 모델이 각기 따로 존재했다. 그러나 Transformer가 소개된 이후로 &lt;strong&gt;&lt;div&gt;self-attention 모듈을 계속 쌓아가는 방식을 이용하여 모델 크기를 키우고, 추가적인 labeling 없이 대규모 텍스트 데이터를 통해 &lt;code&gt;자가지도학습(self-supervised learning)&lt;/code&gt;을 수행하는 방식이 주류&lt;/div&gt;&lt;/strong&gt;가 되었다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;BERT&lt;/code&gt;, &lt;code&gt;GPT-3&lt;/code&gt; 등&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 방식으로 &lt;code&gt;transfer-learning(전이 학습)&lt;/code&gt;을 각 task에 적용하였을 때, 기존의 특화 모델들보다 월등한 성능을 내었다.  따라서 &lt;strong&gt;현재는 NLP task에 있어서 대규모 데이터를 처리하는 방식이 필수적이어서, 적은 GPU 자원으로는 연구하기 어렵게 되었다.&lt;/strong&gt;&lt;/p&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;bag-of-words&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#bag-of-words&quot; aria-label=&quot;bag of words permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Bag-of-Words&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Bag-of-Words&lt;/code&gt;&lt;/strong&gt;란, 자연어처리에 딥러닝 기술이 적용되기 이전에 사용되던 방식으로, &lt;strong&gt;단어 및 문서를 숫자(벡터) 형태로 나타내는 간단한 기법&lt;/strong&gt;이다.&lt;/p&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;과정&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B3%BC%EC%A0%95&quot; aria-label=&quot;과정 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;과정&lt;/h3&gt;&lt;ol&gt;&lt;li&gt;unique word를 포함하는 단어를 추출한다.&lt;/li&gt;&lt;li&gt;각 단어를 &lt;code&gt;categorical variable(범주형 변수)&lt;/code&gt;로 판단하여 원-핫벡터로 나타낸다.&lt;ul&gt;&lt;li&gt;모든 단어 간 유클리드 거리가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msqrt&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msqrt&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sqrt{2}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.04em;vertical-align:-0.13278em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord sqrt&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.90722em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;padding-left:0.833em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.86722em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail&quot; style=&quot;min-width:0.853em;height:1.08em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;1.08em&quot; viewBox=&quot;0 0 400000 1080&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M95,702
c-2.7,0,-7.17,-2.7,-13.5,-8c-5.8,-5.3,-9.5,-10,-9.5,-14
c0,-2,0.3,-3.3,1,-4c1.3,-2.7,23.83,-20.7,67.5,-54
c44.2,-33.3,65.8,-50.3,66.5,-51c1.3,-1.3,3,-2,5,-2c4.7,0,8.7,3.3,12,10
s173,378,173,378c0.7,0,35.3,-71,104,-213c68.7,-142,137.5,-285,206.5,-429
c69,-144,104.5,-217.7,106.5,-221
l0 -0
c5.3,-9.3,12,-14,20,-14
H400000v40H845.2724
s-225.272,467,-225.272,467s-235,486,-235,486c-2.7,4.7,-9,7,-19,7
c-6,0,-10,-1,-12,-3s-194,-422,-194,-422s-65,47,-65,47z
M834 80h400000v40h-400000z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.13278em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다.&lt;/li&gt;&lt;li&gt;모든 단어 간 cosine 유사도가 0이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;문장이나 문서를 원-핫벡터의 합벡터, 즉 Bag-of-words 벡터로 나타낸다.&lt;/li&gt;&lt;li&gt;Bag-of-words를 사용하여 Documents Classification을 수행한다.&lt;/li&gt;&lt;/ol&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;naive-bayes-classifier&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#naive-bayes-classifier&quot; aria-label=&quot;naive bayes classifier permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Naive Bayes Classifier&lt;/h2&gt;&lt;p&gt;Bag-of-words를 사용하여 문서 분류(Documents Classification)를 수행하는 대표적인 방식이다.&lt;/p&gt;&lt;h1 id=&quot;-7&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-7&quot; aria-label=&quot; 7 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;베이지안-룰bayes-rule&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%B2%A0%EC%9D%B4%EC%A7%80%EC%95%88-%EB%A3%B0bayes-rule&quot; aria-label=&quot;베이지안 룰bayes rule permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;베이지안 룰(Bayes&amp;#x27; Rule)&lt;/h3&gt;&lt;p&gt;Document &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 클래스 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;c&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 있다고 할 때,&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;arg max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;arg max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;arg max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
C_{MAP} &amp;amp;= \underset{c\in C}{\argmax}P(c|d)\\
&amp;amp;= \underset{c\in C}{\argmax}\frac{P(d|c)P(c)}{P(d)}\\
&amp;amp;= \underset{c\in C}{\argmax}P(d|c)P(c)\\
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:6.905422999999999em;vertical-align:-3.2027115em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.7027114999999995em&quot;&gt;&lt;span style=&quot;top:-6.2897115em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.427em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.5965705em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.427em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.4904295000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.427em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.2027115em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.7027114999999995em&quot;&gt;&lt;span style=&quot;top:-6.2897115em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.427em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.161229em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.966141em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.5965705em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.427em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.161229em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.966141em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.4904295000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.427em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.161229em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.966141em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.2027115em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;MAP&lt;/code&gt;은 &amp;#x27;Maximum A Posteriori&amp;#x27;, 즉 most likely class를 뜻한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(d)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 document &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 뽑힐 확률인데, 여기서 문서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 이미 정해져있다고 볼 수 있으므로  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(d)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 1이다. 따라서 분모를 제거한다.&lt;/li&gt;&lt;/ul&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;→&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(d|c)P(c) = P(w_1, w_2, \dots, w_n|c)P(c) \rarr p(c)\prod_{w_i\in W}P(w_i|c)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;→&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.4444410000000003em;vertical-align:-1.394436em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.855664em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.02691em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∏&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.394436em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(d|c)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 특정 카테고리 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;c&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 고정 되었을 때 문서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 나타날 확률을 의미하는데, 문서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 곧 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;w_1,\dots,w_n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;까지 n개 단어들이 동시에 나타나는 사건과 같다.&lt;/li&gt;&lt;li&gt;따라서, &lt;strong&gt;&lt;div&gt;각각의 단어들의 출현이 서로 독립이라고 가정할 때&lt;/div&gt;&lt;/strong&gt;, 조건부독립을 활용하여 각 단어가 나타낼 수 있는 확률을 모두 곱한 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;msub&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(c)\prod_{w_i\in W}P(w_i|c)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.14981em;vertical-align:-0.39981em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∏&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.17862099999999992em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.02691em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.39981em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 나타낼 수 있다.&lt;/li&gt;&lt;li&gt;즉, 문서 분류를 위해서는 &lt;strong&gt;각각의 단어들이 나타날 확률&lt;/strong&gt; &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(w_i|c)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 구해야한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 경우, &lt;strong&gt;&lt;div&gt;한번도 출현하지 않은(학습되지 않은) 단어를 분류하는 것은 확률이 0이 되므로 불가능하게 된다.&lt;/div&gt;&lt;/strong&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[NLP 02 - Word Embedding 알아보기]]></title><description><![CDATA[Word Embedding by 주재걸 교수님, BoostCamp AI Tech 4주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/23_word_embedding/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/23_word_embedding/</guid><pubDate>Mon, 15 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;word-embedding&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#word-embedding&quot; aria-label=&quot;word embedding permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Word Embedding&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;단어 임베딩(Word Embedding)&lt;/code&gt;&lt;/strong&gt;이란, &lt;strong&gt;단어를 하나의 벡터로 표현함으로써, n차원의 점으로 나타내는 방식&lt;/strong&gt;이다. 임베딩에서 단어의 의미를 잘 반영했다고 가정할 때, 약 두 단어가 서로 유사한 단어라면 두 점 간의 거리가 작을 것이고, 그렇지 않다면 크게 된다.&lt;/p&gt;&lt;h2 id=&quot;word2vec&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#word2vec&quot; aria-label=&quot;word2vec permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Word2Vec&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Word2Vec&lt;/code&gt;&lt;/strong&gt;은 단어 임베딩을 학습하는 대표적 방법이다.&lt;/p&gt;&lt;p&gt;Word2Vec은 같은 문장 내에 나타난 &lt;strong&gt;인접한 단어들간의 의미가 비슷할 것이라고 가정&lt;/strong&gt;한다.&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;학습-방식&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%95%99%EC%8A%B5-%EB%B0%A9%EC%8B%9D&quot; aria-label=&quot;학습 방식 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;학습 방식&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:47.65625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;Word2Vec&quot; title=&quot;Word2Vec&quot; src=&quot;/static/f11ce1e639d2899a3201ee5bc2d146cd/2bef9/Word2Vec.png&quot; srcSet=&quot;/static/f11ce1e639d2899a3201ee5bc2d146cd/6f3f2/Word2Vec.png 256w,/static/f11ce1e639d2899a3201ee5bc2d146cd/01e7c/Word2Vec.png 512w,/static/f11ce1e639d2899a3201ee5bc2d146cd/2bef9/Word2Vec.png 1024w,/static/f11ce1e639d2899a3201ee5bc2d146cd/71c1d/Word2Vec.png 1536w,/static/f11ce1e639d2899a3201ee5bc2d146cd/a878e/Word2Vec.png 2048w,/static/f11ce1e639d2899a3201ee5bc2d146cd/e3261/Word2Vec.png 2324w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ol&gt;&lt;li&gt;주어진 문장을 Tokenize한다.&lt;/li&gt;&lt;li&gt;unique words를 모아서 Vocabulary를 구성한다.&lt;/li&gt;&lt;li&gt;각 단어를 Vocab size만큼의 dimension을 가지는 원-핫벡터로 표현한다.&lt;/li&gt;&lt;li&gt;Sliding Window 기법을 적용하여, 한 단어를 중심으로 앞 뒤로 나타난 각 word와 함께 입출력 쌍을 구성한다.&lt;ul&gt;&lt;li&gt;e.g. &lt;code&gt;&amp;quot;I study math&amp;quot;&lt;/code&gt; → &lt;code&gt;[(I,study), (study,I), (study,math), (math,study)]&lt;/code&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;원-핫벡터 size만큼의 입/출력 size를 갖는 2-layer nn를 만든다.  이 때, &lt;strong&gt;입력노드와 출력노드 사이의 hidden layer의 노드 수는&lt;/strong&gt; &lt;code&gt;hyperparameter&lt;/code&gt;로, 워드 임베딩을 수행할 좌표공간의 차원수로 설정한다.&lt;ul&gt;&lt;li&gt;따라서 weight는 (#hidden nodes, #input nodes), (#output nodes, #hidden nodes)의 shape가 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;입력값을 넣어 linear 변환을 하고, softmax 함수를 적용시켜 확률분포 벡터(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\hat{y}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.19444em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)를 만든다.&lt;ul&gt;&lt;li&gt;이 때 입력값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 원-핫벡터이므로, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Wx&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; &lt;strong&gt;내적 시에 해당 단어와 관련있는 단어들만 1이 곱해져 반영되고, 그렇지 않은 단어들은 0이 되어 반영되지 않는다.&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;따라서 실제로 구현할 때는 이 레이어를 &lt;strong&gt;&lt;code&gt;Embedding Layer&lt;/code&gt;&lt;/strong&gt;라고 부르고, &lt;strong&gt;일일이 행렬곱하는 대신 반영될 column만 바로 뽑아온다&lt;/strong&gt;(위의 사진에서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 파란 컬럼)&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;확률분포벡터와 ground truth(정답 레이블)를 비교하여 거리가 가까워지도록 &lt;strong&gt;&lt;code&gt;softmax loss&lt;/code&gt;&lt;/strong&gt;를 적용시킨다.&lt;ul&gt;&lt;li&gt;이 때, ground truth값과 비슷해지도록 하려면, 이론적으로는 softmax 함수의 logit값, 즉 linear 변환을 모두 통과한 값이  정답 클래스만 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∞&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\infin&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∞&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 되고, 나머지는 모두 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∞&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;-\infin&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.66666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∞&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 되어야한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;파라미터를 모두 학습시키고 난 뒤, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;과 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 임베딩 벡터 둘 중 아무것이나 Word2Vec의 모델로 사용해도 된다. &lt;strong&gt;&lt;div&gt;이는 Word2Vec이 기본적으로 단어간의 관계도를 측정하는 방식이기 때문에, 순서 상관없이 단어와 단어 간의 weight vector만 있으면 되기 때문&lt;/div&gt;&lt;/strong&gt;이다. 그러나, 컨벤션으로는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 많이 사용한다. 또, output이 나오면 &lt;strong&gt;&lt;code&gt;PCA(Principal Component Analysis)&lt;/code&gt;&lt;/strong&gt;를 통해서 2차원으로 차원축소하여 scatter plot 형태로 시각화하곤 한다.&lt;/p&gt;&lt;p&gt;다른 예시가 보고싶다면, 다음 사이트를 참조해보자. 워드 임베딩 과정을 시각적으로 보여주는 사이트이다.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://ronxin.github.io/wevi/&quot;&gt;wevi&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;성질property&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%84%B1%EC%A7%88property&quot; aria-label=&quot;성질property permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;성질(property)&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:33.203125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAYAAAAIy204AAAACXBIWXMAAAsSAAALEgHS3X78AAAA6ElEQVQoz3VRCW7DMAzL//85IKjj+Ijvi5O8uejWVAAhWLJFUt7wG2OMmXsfiLnhaxe4nIe2F0pt4G4uFcdDoPVO4Mp4vluxcWGBw4cwhyhtcBpH5wh+a5RCKQXNWmTnPg9c6n4aA7V11FpgXYRUBq3Vqa76gHpZXLFAHifV+x9nz4HM+j8CqXIhkdWOnAtiytP2KU+EVCYBK3x1t7A5kp9zRoxxWuLMZ1YphICUcuI4DiQafKfqzXKgvSnakfcexhhY2hMTceYek2itkVMionqLud/WsO37TszpIyNf6vSrSz2D79+Be9+eBCUatDzKXgAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;Word2Vec_property&quot; title=&quot;Word2Vec_property&quot; src=&quot;/static/9ebdf79d76e0050fd08cdc0df0da471e/2bef9/Word2Vec_property.png&quot; srcSet=&quot;/static/9ebdf79d76e0050fd08cdc0df0da471e/6f3f2/Word2Vec_property.png 256w,/static/9ebdf79d76e0050fd08cdc0df0da471e/01e7c/Word2Vec_property.png 512w,/static/9ebdf79d76e0050fd08cdc0df0da471e/2bef9/Word2Vec_property.png 1024w,/static/9ebdf79d76e0050fd08cdc0df0da471e/5a3c9/Word2Vec_property.png 1169w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;각 단어 벡터, 또는 공간상의 벡터 포인트는 단어간의 관계를 나타낸다. &lt;strong&gt;같은 관계를 가진다면 같은 형태의 vector로 표시&lt;/strong&gt;된다.&lt;/p&gt;&lt;p&gt;우리말로 Word2Vec을 학습시켜놓은 사이트를 참고해보자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://word2vec.kr/search/&quot;&gt;Korean Word2Vec&lt;/a&gt;&lt;/p&gt;&lt;p&gt;또, Word2Vec을 이용하여  여러 단어들 중 가장 상이한 단어를 골라내는 &lt;code&gt;Word Intrusion Dectection&lt;/code&gt;을 수행할 수도 있다. 여러 단어가 주어지면, 한 단어를 중심으로 다른 단어들의 유클리드 거리를 계산하여 평균낸다. 모든 단어에 대해 이를 수행하였을 때, 유클리드 거리 평균이 가장 먼 단어, 즉 가장 관계가 없는 단어를 찾아 솎아낼 수 있다.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://github.com/dhammack/Word2VecExample&quot;&gt;dhammack/Word2VecExample&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;적용&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A0%81%EC%9A%A9&quot; aria-label=&quot;적용 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;적용&lt;/h3&gt;&lt;p&gt;Word2Vec은 자연어 처리 대부분의 분야에서 성능 개선에 사용되고 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;단어 유사도&lt;/li&gt;&lt;li&gt;기계 번역(Machine Translation)&lt;/li&gt;&lt;li&gt;형태소 분석(Part-of-speech(POS) tagging)&lt;/li&gt;&lt;li&gt;개체명 인식(Named entity recognition,NER)&lt;/li&gt;&lt;li&gt;감정 분석&lt;/li&gt;&lt;li&gt;군집화&lt;/li&gt;&lt;li&gt;어휘 의미 구성(semantic lexicon building)&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;또, 이미지를 설명하는 캡션을 생성하는 &lt;code&gt;Image Captioning&lt;/code&gt; 분야에도 사용된다.&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h1 id=&quot;glove&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#glove&quot; aria-label=&quot;glove permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GloVe&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Glove : Gloval Vectors for Word Presentation&lt;/code&gt;&lt;/strong&gt;는 Word2Vec와 함께 워드 임베딩의 가장 대표적인 기법 중 하나이다.&lt;/p&gt;&lt;p&gt;Word2Vec과 다른 GloVe의 차이점은 &lt;strong&gt;&lt;div&gt;학습과정에서 두 단어가 한 윈도우 내에서 몇번이나 동시 등장했는가&lt;/div&gt;&lt;/strong&gt;를 사전에 기록해둔다는 것이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;J&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mfrac&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/munderover&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;msubsup&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mstyle&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msub&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mstyle mathcolor=&quot;Green&quot;&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;J(\theta) = \frac{1}{2}\sum^W_{i,j=1}f(P_{i,j})(\textcolor{blue}{u^T_i}\textcolor{red}{v_j} - log\textcolor{Green}{P_{ij}})^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.09618em&quot;&gt;J&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.2421130000000007em;vertical-align:-1.4137769999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.8283360000000006em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.4137769999999998em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em;color:blue&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:red&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em;color:red&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.150216em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:Green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:Green&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:Green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:Green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:Green&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em;color:Green&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;msubsup&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msubsup&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{blue}{u_i^T}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0999949999999998em;vertical-align:-0.258664em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-2.441336em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em;color:blue&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.258664em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 입력 word의 임베딩 벡터&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msub&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{red}{v_j}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em;color:red&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em;color:red&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 출력 word의 임베딩 벡터&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mstyle mathcolor=&quot;Green&quot;&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textcolor{Green}{P_{ij}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.969438em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:Green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:Green&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:Green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:Green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:Green&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em;color:Green&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 두 단어가 한 윈도우 내에서 동시 출현한 count&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;위의 비용함수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;J&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;J(\theta)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.09618em&quot;&gt;J&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서 cost를 최소화하려면, &lt;strong&gt;입력/출력 단어의 내적값(관계도)이 동시 출현한 count와 비슷해지도록 조정&lt;/strong&gt;해야한다.&lt;/p&gt;&lt;p&gt;Word2Vec은 특정 입출력 단어쌍이 자주 등장할수록 여러번에 걸쳐 학습되게 하여, 학습이 빈번할수록 두 임베딩 벡터간의 내적값이 더 커지도록 만들었다.&lt;/p&gt;&lt;p&gt;반면, GloVe는 &lt;strong&gt;단어쌍이 동시에 등장한 횟수를 미리 계산하고, 이 count를 학습과정에서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;logP_{i,j}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.980548em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라는 ground truth로 사용함으로써 계산과정의 중복을 제거&lt;/strong&gt;하였다. 따라서 &lt;div&gt;학습시간이 Word2Vec보다 더 빠르며, 더 적은 데이터에도 잘 동작&lt;/div&gt;한다.&lt;/p&gt;&lt;p&gt;또한, GloVe는 추천 시스템에 사용되는 핵심 로직인 &lt;code&gt;co-occurence matrix&lt;/code&gt;의 &lt;code&gt;low rank matrix factorization&lt;/code&gt;의 task로도 이해할 수 있다. &lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://nlp.stanford.edu/projects/glove/&quot;&gt;GloVe: Global Vectors for Word Representation&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[딥러닝 베이직 - 08.Genarative Model 기초]]></title><description><![CDATA[Generative Models by 최성준 교수님, BoostCamp AI Tech 3주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/21_generative_models/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/21_generative_models/</guid><pubDate>Fri, 05 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;generative-models&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#generative-models&quot; aria-label=&quot;generative models permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Generative Models&lt;/h1&gt;&lt;p&gt;시작하기 앞서서, Generative Models를 다룬 스탠퍼드대 CS236 강의 블로그를 소개한다.&lt;/p&gt;&lt;p&gt; 필요할 때 syllabus와 슬라이드를 참고하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://deepgenerativemodels.github.io/&quot;&gt;Deep Generative Models&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;우리는 &lt;strong&gt;&lt;code&gt;생성 모델(Generative Model)&lt;/code&gt;&lt;/strong&gt;을 생각하면 대부분 &lt;strong&gt;&lt;code&gt;적대적 생성모델(GAN)&lt;/code&gt;&lt;/strong&gt;을 떠올린다. 그러나 그럴듯한 이미지나 문장을 만드는 것이 생성 모델의 전부는 아니다.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;생성모델의-학습방법&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%83%9D%EC%84%B1%EB%AA%A8%EB%8D%B8%EC%9D%98-%ED%95%99%EC%8A%B5%EB%B0%A9%EB%B2%95&quot; aria-label=&quot;생성모델의 학습방법 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;생성모델의 학습방법&lt;/h2&gt;&lt;p&gt;개(dog)의 이미지가 여러 장 주어졌다고 하자.&lt;/p&gt;&lt;p&gt;생성모델의 &lt;code&gt;확률분포(probability distribution)&lt;/code&gt; &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 학습함으로써 할 수 있는 것은 다음과 같다. &lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Generation&lt;/code&gt;&lt;/strong&gt; : 확률분포 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 유사한 어떤 새로운 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_{new}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 만들어서 (&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_{new} \sim p(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;), 개와 유사한 이미지를 만들어낼 수 있다. (sampling)&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Density estimation&lt;/code&gt;&lt;/strong&gt; : 어떤 이미지 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 개인지 아닌지 판별해낼 수 있다. &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 높게 나온다면 개이고, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 낮게 나온다면 개가 아니다. (anomaly detection)&lt;ul&gt;&lt;li&gt;엄밀한 의미의 생성 모델은 &lt;strong&gt;&lt;code&gt;Discriminator&lt;/code&gt;&lt;/strong&gt; 모델까지 포함하고 있다. 어떤 이미지가 특정 레이블에 속하는 지 아닌지 판단할 수 있어야 한다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;div&gt;이처럼 Discriminator를 포함한 생성 모델을 &lt;code&gt;명시적 모델(Explicit Model)&lt;/code&gt;이라고 하며, Generation만 할 수 있는 모델을 &lt;code&gt;암시적 모델(Implicit Model)&lt;/code&gt;이라고 한다.&lt;/div&gt;&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;Unsupervised representation learning&lt;/code&gt; : 이미지들 사이에서 feature(특징,공통점)을 찾아낼 수 있다.(feature learning)&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;그렇다면 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 무엇일까? &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 집어넣었을 때 나오는 값일수도 있고, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 샘플링할 수 있는 어떤 모델일수도 있다. 이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 알기 위해서는 먼저 확률에 대한 간단한 선행지식이 필요하다.&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;basic-discrete-distributions&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#basic-discrete-distributions&quot; aria-label=&quot;basic discrete distributions permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Basic Discrete Distributions&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;베르누이 분포(Bernoulli distribution)&lt;/code&gt;&lt;/strong&gt; : 앞, 또는 뒤만 나오는 동전 던지기와 같다.(0,1)&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;D&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;= [앞면(Head), 뒷면(Tails)]&lt;/li&gt;&lt;li&gt;앞면이 나올 확률이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(X=Heads)=p&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 라면, 뒷면이 나올 확률은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(X=Tails) = 1-p&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다.&lt;/li&gt;&lt;li&gt;이를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X \sim Ber(p)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 표현한다.&lt;/li&gt;&lt;li&gt;파라미터는 1개(p 하나)&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;카테고리 분포(Categorical Distribution)&lt;/code&gt;&lt;/strong&gt; : &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;m&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 개의 면을 가지는 주사위를 던지는 것과 같다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;D&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; = [1, 2, ... , m]&lt;/li&gt;&lt;li&gt;주사위를 던져 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 가 나올 확률을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;Y&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(Y=i)=p_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;Y&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 한다면, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sum^m_{i=1}p_i=1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.104002em;vertical-align:-0.29971000000000003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.804292em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2029em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29971000000000003em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다.&lt;/li&gt;&lt;li&gt;이를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Y&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Y \sim Cat(p_1, \cdots,p_m)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;Y&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이라고 한다.&lt;/li&gt;&lt;li&gt;파라미터는 m-1개&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;예시&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%98%88%EC%8B%9C&quot; aria-label=&quot;예시 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;예시&lt;/h3&gt;&lt;p&gt;RGB 픽셀 하나의 &lt;code&gt;결합분포(joint distribution)&lt;/code&gt;를 모델링한다고 생각하자.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(r,g,b) \sim p(R,G,B)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.00773em&quot;&gt;R&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;가능한 색상의 종류(경우의 수) :  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;256&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;256&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;256&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;256\times 256\times 256&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;6&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;6&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;6&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 개&lt;/li&gt;&lt;li&gt;필요한 파라미터의 개수 : &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;255&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;255&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;255&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;255\times 255\times 255&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;단 하나의 RGB 픽셀을 표현하기 위한 &lt;strong&gt;파라미터 숫자가 굉장히 많다&lt;/strong&gt;.&lt;/p&gt;&lt;p&gt;이번엔 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개의 binary pixels(하나의 binary image) &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_1,\dots,X_n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 있다고 치자.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;가능한 경우의 수 : &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msup&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;2\times 2\times \cdots \times 2 = 2^n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.66666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.664392em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.664392em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 개&lt;/li&gt;&lt;li&gt;필요한 파라미터의 개수 : &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msup&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;2^n - 1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.747722em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.664392em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이처럼, 바이너리 이미지를 나타내는 데에도 &lt;strong&gt;너무 많은 파라미터가 필요&lt;/strong&gt;하다.&lt;/p&gt;&lt;p&gt;그러나 파라미터의 개수가 늘어나면 모델의 학습은 일반적으로 잘 되지 않는다. 이 파라미터를 줄일 방법은 없을까?&lt;/p&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;structure-through-independence&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#structure-through-independence&quot; aria-label=&quot;structure through independence permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Structure Through Independence&lt;/h3&gt;&lt;p&gt;위의 binary image 사례에서, 모든 픽셀 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_1,\dots,X_n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 각각 &lt;strong&gt;독립적이라고 가정(independance assumption)&lt;/strong&gt;하면, 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_1, \dots, x_n)= p(x_1)p(x_2)\cdots p(x_n)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;가능한 경우의 수 : &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;2^n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.664392em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.664392em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 개&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_1,\dots,x_n)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 구하기 위해 필요한 파라미터의 개수 : &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;  개&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;동일하게 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;2^n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.664392em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.664392em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개의 경우의 수를 나타낼 수 있지만, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 개의 파라미터만 있으면 된다.&lt;/p&gt;&lt;p&gt;당연히 말이 안되는 가정이긴 하다. 이미지사진이므로 각 픽셀사이의 분포는 아무래도 인접할수록 비슷할 확률이 높은데, 이를 완전히 무시한 가정이기 때문에, 적당한 분포를 모델링하기에는 좋지 않다.&lt;/p&gt;&lt;h1 id=&quot;-7&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-7&quot; aria-label=&quot; 7 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;conditional-independence&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#conditional-independence&quot; aria-label=&quot;conditional independence permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Conditional Independence&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;기존의 Fully Dependant 모델링과 Independent 모델링의 중간점으로 타협한 것&lt;/strong&gt;이 &lt;strong&gt;&lt;code&gt;Conditional Independence&lt;/code&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;h1 id=&quot;-8&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-8&quot; aria-label=&quot; 8 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;Conditional Indepence는 다음과 같은 3가지 핵심 룰로 동작한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;연쇄법칙(Chain rule)&lt;/code&gt;&lt;/strong&gt;&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_1,\dots,x_n) = p(x_1)p(x_2|x_1)p(x_3|x_1,x_2)\cdots p(x_n|x_1,\cdots,x_{n-1})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;베이즈 정리(Bayes&amp;#x27; rule)&lt;/code&gt;&lt;/strong&gt;&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x|y) = \frac{p(x,y)}{p(y)}=\frac{p(y|x)p(x)}{p(y)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.363em;vertical-align:-0.936em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.363em;vertical-align:-0.936em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;조건부독립&lt;/code&gt;&lt;/strong&gt;&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mtext&gt;if &lt;/mtext&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;⊥&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mtext&gt; then &lt;/mtext&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textrm{if} \ x \perp y |z, \  \textrm{then} \ p(x|y,z) = p(x|z)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;if&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;⊥&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;then&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 주어졌을 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 독립적이라면, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y,z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 주어졌을 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 일어날 확률은 그냥 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;만 주어지더라도 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 가 일어날 확률과 같다. 즉 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 상관이 없다.&lt;/li&gt;&lt;li&gt;이것으로 연쇄법칙의 conditional 부분을 날릴 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-9&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-9&quot; aria-label=&quot; 9 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;조건부 독립이 잘 이해되지 않는다면, 적절한 예시를 들어 잘 설명해놓은 다음 글을 참고한다.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://actruce.com/conditional-independence/&quot;&gt;조건부 독립 ( Conditional Independence )&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-10&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-10&quot; aria-label=&quot; 10 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;그럼 binary 이미지 예제에서 연쇄법칙을 이용한다고 가정해보자.&lt;/p&gt;&lt;p&gt;필요한 파라미터의 개수는 몇개일까?&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_1)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 1개&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_2|x_1)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 2개 (하나는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_2|x_1)=0&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, 나머지 하나는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_2|x_1)=1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_3|x_1,x_2)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 4개 (입력 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_1,x_2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 모두 고려한다)&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;따라서 , 최종적으로 필요한 파라미터의 개수는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msup&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msup&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msup&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msup&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;1+2+2^2+\cdots+2^{n-1} = 2^n-1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.897438em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.66666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8141079999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.747722em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.664392em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로, Fully Dependant 모델과 같다.&lt;/p&gt;&lt;h1 id=&quot;-11&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-11&quot; aria-label=&quot; 11 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이제, &lt;code&gt;Markov assumption&lt;/code&gt;을 가정해보자. 즉, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_{i+1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.891661em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_{i}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에만 dependant하고, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_1,\dots,X_{i-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.891661em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 까지에는 independent하다고 가정한다.(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;⊥&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_{i+1}\perp X_1,\dots, X_{i-1}|X_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.902771em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;⊥&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_1,\dots,x_n) = p(x_1)p(x_2|x_1)p(x_3|x_2)\cdots p(x_n|x_{n-1})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이 경우 파라미터 개수는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;2n-1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다.&lt;/p&gt;&lt;p&gt;따라서, Markov assumption을 적용시킴으로써 &lt;strong&gt;파라미터의 개수를 지수차원에서 끌어내릴 수 있다&lt;/strong&gt;.&lt;/p&gt;&lt;h1 id=&quot;-12&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-12&quot; aria-label=&quot; 12 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Autoregressive Model&lt;/code&gt;&lt;/strong&gt;은 이 conditional independance를 잘 활용한 모델이다.&lt;/p&gt;&lt;h1 id=&quot;-13&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-13&quot; aria-label=&quot; 13 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;auto-regressive-model&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#auto-regressive-model&quot; aria-label=&quot;auto regressive model permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Auto-regressive Model&lt;/h3&gt;&lt;p&gt;28x28크기의 바이너리 이미지(픽셀들의 모음)이 있다고 하자.&lt;/p&gt;&lt;p&gt;우리는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;{&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;}&lt;/mo&gt;&lt;mn&gt;784&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x \in \{0,1\}^{784}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.5782em;vertical-align:-0.0391em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.064108em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;7&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;8&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;인 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;784&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x)=p(x_1,\dots,x_{784})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;7&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;8&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 구해야한다. 이 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 어떻게 파라미터로 표현할 수 있을까?&lt;/p&gt;&lt;h1 id=&quot;-14&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-14&quot; aria-label=&quot; 14 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;ul&gt;&lt;li&gt;연쇄법칙을 이용해 결합분포를 나눈다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mn&gt;784&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_{1:784})=p(x_1)p(x_2|x_1)p(x_3|x_{1:2})\cdots&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;7&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;8&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;이것을 &lt;strong&gt;&lt;code&gt;AR모델(Autoregressive Model, 자기회귀모델)&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;ul&gt;&lt;li&gt;AR모델은 &lt;strong&gt;하나의 정보가 이전 정보에 dependant한 것&lt;/strong&gt;을 의미한다. 바꿔말하면, Markov assumption처럼 직전 정보에만 dependant한것도 AR모델이고, 거꾸로 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_1\dots,x_{i-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.638891em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;까지에 모두 dependant한 것도 AR모델이다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;이전의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개 정보들을 고려하는 모델&lt;/strong&gt;을 &lt;strong&gt;&lt;code&gt;AR(n) 모델&lt;/code&gt;&lt;/strong&gt;이라고 부른다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;이 때, 주의할 점은 바이너리 픽셀들을 1부터 784까지 순서를 매겼듯이 &lt;strong&gt;랜덤한 variable들에 각각 순서를 매겨야 한다&lt;/strong&gt;는 것이다.&lt;ul&gt;&lt;li&gt;이 순서를 어떻게 정하느냐에 따라 모델의 성능이 달라질수도 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-15&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-15&quot; aria-label=&quot; 15 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;어떤 식으로 Coditional Indepedency를 주는가에 따라 연쇄법칙을 이용해 결합분포를 나누는 방식에 차이가 생기므로, 결과적으로 AR모델의 structure가 달라진다.&lt;/p&gt;&lt;h1 id=&quot;-16&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-16&quot; aria-label=&quot; 16 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;생성모델의-종류&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%83%9D%EC%84%B1%EB%AA%A8%EB%8D%B8%EC%9D%98-%EC%A2%85%EB%A5%98&quot; aria-label=&quot;생성모델의 종류 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;생성모델의 종류&lt;/h2&gt;&lt;h3 id=&quot;nade--neural-autoregressive-density-estimator&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#nade--neural-autoregressive-density-estimator&quot; aria-label=&quot;nade  neural autoregressive density estimator permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;NADE : Neural Autoregressive Density Estimator&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:750px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:37.5%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;nade&quot; title=&quot;nade&quot; src=&quot;/static/8091191a89534c015807b1fd896bcc27/1d69c/nade.png&quot; srcSet=&quot;/static/8091191a89534c015807b1fd896bcc27/6f3f2/nade.png 256w,/static/8091191a89534c015807b1fd896bcc27/01e7c/nade.png 512w,/static/8091191a89534c015807b1fd896bcc27/1d69c/nade.png 750w&quot; sizes=&quot;(max-width: 750px) 100vw, 750px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;α&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mtext&gt; where &lt;/mtext&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo&gt;&amp;lt;&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_i|x_{1:i-1}) = \sigma(\alpha_ih_i+b_i) \ \textrm{where}\ h_i=\sigma(W_{&amp;lt;i}x_{1:i-1}+c)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.0037em&quot;&gt;α&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.0037em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;where&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mrel mtight&quot;&gt;&amp;lt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.17737em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째 픽셀 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 그 이전까지의 모든 픽셀들에 대해 dependant하다. 픽셀의 순서가 뒤로 갈수록 받는 입력(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_{1:i-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.638891em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)의 수가 많아지므로, &lt;strong&gt;weight의 길이가 가변적&lt;/strong&gt;이다. 그 이외에는 AR모델과 동일하다.&lt;/p&gt;&lt;h1 id=&quot;-17&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-17&quot; aria-label=&quot; 17 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;NADE&lt;/code&gt;&lt;/strong&gt;는 &lt;code&gt;explicit 모델&lt;/code&gt;로, &lt;strong&gt;주어진 입력값의 확률(density)를 계산할 수 있다.&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;784 픽셀의 바이너리 이미지라고 가정하면, 각 조건부 확률 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_i|x_{1:i-1})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 독립적으로 계산될 때 결합분포를 다음과 같이 계산할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;784&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;784&lt;/mn&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mn&gt;783&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_1,\dots,x_{784})=p(x_1)p(x_2|x_1)\cdots p(x_{784}|x_{1:783})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;7&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;8&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;7&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;8&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;7&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;8&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;각각의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x_i|x_{1:i-1})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 차례차례 계산해 대입하면, 전체 확률을 알 수 있게 된다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;NADE는 이 확률을 토대로 해당 이미지를 판별하는 &lt;strong&gt;discriminator 역할을 수행할 수 있다&lt;/strong&gt;. 그래서 논문 제목에 &lt;code&gt;Density Estimator&lt;/code&gt;라는 단어가 붙었다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&amp;quot;Density Estimator&amp;quot;라는 단어는 explicit model을 표현할 때 많이 사용되는 단어이다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-18&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-18&quot; aria-label=&quot; 18 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;위의 예제에서는 이산변수인 바이너리 픽셀이기때문에 &lt;code&gt;sigmoid&lt;/code&gt;를 통과시킬 수 있었지만, 임의의 연속변수를 모델링할 때는 &lt;code&gt;가우시안 혼합(Gaussian mixture) 모델&lt;/code&gt;을 이용할 수 있다.&lt;/p&gt;&lt;h3 id=&quot;pixel-rnn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#pixel-rnn&quot; aria-label=&quot;pixel rnn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Pixel RNN&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Pixel RNN&lt;/code&gt;&lt;/strong&gt;은 AR 모델을 만들기 위해 RNN을 활용하는 방법이다.&lt;/p&gt;&lt;h1 id=&quot;-19&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-19&quot; aria-label=&quot; 19 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n\times n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.66666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 사이즈의 RGB 모델이 있다면,&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;msup&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/munderover&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mo&gt;&amp;lt;&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;mstyle mathcolor=&quot;green&quot;&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mo&gt;&amp;lt;&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mo&gt;&amp;lt;&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p(x) = \prod^{n^2}_{i=1}\textcolor{red}{p(x_{i,R}|x_{&amp;lt;i})}\textcolor{green}{p(x_{i,G}|x_{&amp;lt;i},x_{i,R})}\textcolor{blue}{p(x_{i,B}|x_{&amp;lt;i},x_{i,R},x_{i,G})}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.2515940000000003em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.9739250000000004em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∏&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913142857142857em&quot;&gt;&lt;span style=&quot;top:-2.931em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.328331em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot; style=&quot;color:red&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.00773em;color:red&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:red&quot;&gt;&amp;lt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.17737em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:red&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:green&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.328331em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot; style=&quot;color:green&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:green&quot;&gt;&amp;lt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.17737em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:green&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:green&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.328331em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:green&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:green&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot; style=&quot;color:green&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.00773em;color:green&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:green&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:blue&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.328331em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot; style=&quot;color:blue&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05017em;color:blue&quot;&gt;B&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:blue&quot;&gt;&amp;lt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.17737em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:blue&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.328331em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot; style=&quot;color:blue&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.00773em;color:blue&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:blue&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.328331em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:blue&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot; style=&quot;color:blue&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot; style=&quot;color:blue&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;위와 같이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째 확률을 R,G,B 순으로 구해주고, 앞의 색상 확률들을 조건부로 넣어준다.&lt;/p&gt;&lt;h1 id=&quot;-20&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-20&quot; aria-label=&quot; 20 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;기존 모델은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;1,\dots,i-1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.85396em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째 입력을 고려하기 위해  AR 모델을 &lt;strong&gt;전연결계층(dense layer)을 가진 신경망&lt;/strong&gt;으로 만들었다면,  &lt;strong&gt;&lt;code&gt;Pixel RNN&lt;/code&gt;&lt;/strong&gt;은 &lt;strong&gt;RNN을 이용한다&lt;/strong&gt;는 차이점이 있다.&lt;/p&gt;&lt;h1 id=&quot;-21&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-21&quot; aria-label=&quot; 21 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:51.171875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;pixel_rnn&quot; title=&quot;pixel_rnn&quot; src=&quot;/static/784b1bd43badc4668360095494b5d8da/2bef9/pixel_rnn.png&quot; srcSet=&quot;/static/784b1bd43badc4668360095494b5d8da/6f3f2/pixel_rnn.png 256w,/static/784b1bd43badc4668360095494b5d8da/01e7c/pixel_rnn.png 512w,/static/784b1bd43badc4668360095494b5d8da/2bef9/pixel_rnn.png 1024w,/static/784b1bd43badc4668360095494b5d8da/71c1d/pixel_rnn.png 1536w,/static/784b1bd43badc4668360095494b5d8da/08c33/pixel_rnn.png 1570w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-22&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-22&quot; aria-label=&quot; 22 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이 때 입력의 순서(ordering)를 어떻게 하느냐에 따라 두가지 방법으로 나뉜다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Row LSTM - &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째 픽셀을 만들 때 위쪽 row와 바로 직전 픽셀을 활용한다.&lt;/li&gt;&lt;li&gt;Diagonal BiLSTM - &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째 픽셀을 만들 때 이전의 모든 픽셀들을 활용한다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-23&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-23&quot; aria-label=&quot; 23 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;Pixel RNN의 정확한 동작 방식에 대해서는 다음 블로그를 참고하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;http://ai-hub.kr/post/98/&quot;&gt;PR-024: Pixel Recurrent Neural Network&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-24&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-24&quot; aria-label=&quot; 24 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;latent-variable-models&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#latent-variable-models&quot; aria-label=&quot;latent variable models permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Latent Variable Models&lt;/h2&gt;&lt;p&gt;이번에는 좀 더 실용적으로 많이 사용되는(practical) 생성모델들에 대해 알아보자.&lt;/p&gt;&lt;h1 id=&quot;-25&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-25&quot; aria-label=&quot; 25 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;variational-auto-encoder&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#variational-auto-encoder&quot; aria-label=&quot;variational auto encoder permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Variational Auto-encoder&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://dare.uva.nl/search?identifier=8e55e07f-e4be-458f-a929-2f9bc2d169e8&quot;&gt;Digital Academic Repository - University of Amsterdam&lt;/a&gt;&lt;/p&gt;&lt;div&gt;&lt;p&gt;Variational Auto-encoder와 달리, 그냥 Auto-encoder는 생성 모델이 아니다. Variational auto-encoder와 그냥 auto-encoder에 어떤 차이점이 있는지에 유의하면서 살펴보도록 하자.&lt;/p&gt;&lt;/div&gt;&lt;h1 id=&quot;-26&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-26&quot; aria-label=&quot; 26 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Variational Inference(VI, 변분추론)&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;div&gt;VI의 목적은 &lt;code&gt;사후분포(posterior distribution)&lt;/code&gt;과 비슷한(근사하는) &lt;code&gt;variational distribution&lt;/code&gt;을 찾는(최적화하는)것이다.&lt;/div&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;사후분포&lt;/code&gt; : &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_{\theta}(z|x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;관측값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 주어졌을 때 특정확률변수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 분포. 이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 를 &lt;strong&gt;&lt;code&gt;잠재벡터(latent)&lt;/code&gt;&lt;/strong&gt;라고 부른다.&lt;/li&gt;&lt;li&gt;이를 거꾸로 뒤집은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_\theta(x|z)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 를 &lt;code&gt;Likelihood(가능도,우도)&lt;/code&gt;라고 부른다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;variational distribution&lt;/code&gt;: &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;q_\theta(z|x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;div&gt;일반적으로 사후분포를 정확히 구하는 것은 거의 불가능에 가까우므로, variational distribution을 구해서 사후분포에 근사하도록 최적화시킨다.&lt;/div&gt;&lt;/li&gt;&lt;li&gt;&lt;div&gt;실제 사후분포와의 `쿨백 라이블러 발산(KL Divergence)`이 최소화되는 varitional distribution을 찾는다.&lt;/div&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-27&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-27&quot; aria-label=&quot; 27 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:45.3125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAJCAIAAAC9o5sfAAAACXBIWXMAABYlAAAWJQFJUiTwAAAAxUlEQVQoz41R7Q6CMAzk/R9RQAfqGLDJtjD53OolKBqiwUt/XNJe22sj+gbvvZRSCME5b5qGfiBaWQjvWPRJeuS8mOd5X7zBNE1xkjCWgeyLK0UZJ3alvKCL6M+FuxlvrSnL0lq7I27vpDSlZzpdSFQDF66SgzFGa922bVjM/Ll2CD6OD7z4z/NyKo/wiACrjOUYC/HXsU8xcrjt5lVYVesGhvGq4YVxHFeOphFydV0rpT7XQzvUYXjXdejS9/2qWTiyzrkH5VwPP9jEDAEAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;vae&quot; title=&quot;vae&quot; src=&quot;/static/eafe30704bd43f0e214a205f20b085f2/2bef9/vae.png&quot; srcSet=&quot;/static/eafe30704bd43f0e214a205f20b085f2/6f3f2/vae.png 256w,/static/eafe30704bd43f0e214a205f20b085f2/01e7c/vae.png 512w,/static/eafe30704bd43f0e214a205f20b085f2/2bef9/vae.png 1024w,/static/eafe30704bd43f0e214a205f20b085f2/71c1d/vae.png 1536w,/static/eafe30704bd43f0e214a205f20b085f2/a878e/vae.png 2048w,/static/eafe30704bd43f0e214a205f20b085f2/fc477/vae.png 2426w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-28&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-28&quot; aria-label=&quot; 28 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;그런데, 사후분포가 어떤지도 모르면서 사후분포에 근사하는 값을 찾겠다는 것은 어불성설이 아닌가? 그래서 사용하는것이 &lt;strong&gt;&lt;code&gt;ELBO(Evidence Lower BOund) 방법&lt;/code&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;h1 id=&quot;-29&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-29&quot; aria-label=&quot; 29 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;ln&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo fence=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;ln&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo fence=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;ln&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo fence=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;ln&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo fence=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;ln&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;munder&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo fence=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;ln&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mo&gt;↑&lt;/mo&gt;&lt;/mstyle&gt;&lt;/munder&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;munder&gt;&lt;munder&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mo&gt;↓&lt;/mo&gt;&lt;/mstyle&gt;&lt;/munder&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
p_\theta(D) &amp;amp;= \mathbb{E}_{q_\phi(z|x)}[\ln\ p_\theta(x)]\\
&amp;amp;= \mathbb{E}_{q_\phi(z|x)}\bigg[\ln\ \frac{p_\theta(x,z)}{p_\theta(z|x)}\bigg]\\
&amp;amp;= \mathbb{E}_{q_\phi(z|x)}\bigg[\ln\ \frac{p_\theta(x,z)q_{\theta}(z|x)}{q_{\theta}(z|x)p_\theta(z|x)}\bigg]\\
&amp;amp;= \mathbb{E}_{q_\phi(z|x)}\bigg[\ln\ \frac{p_\theta(x,z)}{q_\theta(z|x)}\bigg] +\mathbb{E}_{q_\phi(z|x)}\bigg[\ln\ \frac{q_\theta(x|z)}{p_\theta(z|x)}\bigg]\\
&amp;amp;= \underbrace{\mathbb{E}_{q_\phi(z|x)}\bigg[\ln\ \frac{p_\theta(x,z)}{q_\theta(z|x)}\bigg]}_{\textcolor{blue}{ELBO\uparrow}} +\underbrace{D_{KL}(q_{\theta}(z|x)\Vert p_\theta(z|x)}_{\textcolor{red}{Objective\downarrow}}\\
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:13.793616em;vertical-align:-6.646808em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:7.146808em&quot;&gt;&lt;span style=&quot;top:-9.756808em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-7.623528em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.923498em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.223468000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:0.4765619999999995em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:6.646808em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:7.146808em&quot;&gt;&lt;span style=&quot;top:-9.756808em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:-0.03588em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29011428571428566em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.38327999999999984em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;ln&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-7.623528em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:-0.03588em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29011428571428566em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.38327999999999984em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;[&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;ln&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.923498em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:-0.03588em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29011428571428566em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.38327999999999984em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;[&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;ln&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.223468000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:-0.03588em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29011428571428566em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.38327999999999984em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;[&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;ln&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:-0.03588em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29011428571428566em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.38327999999999984em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;[&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;ln&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:0.4765619999999995em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.45em&quot;&gt;&lt;span style=&quot;top:-1.165862em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em;color:blue&quot;&gt;E&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05017em;color:blue&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em;color:blue&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:blue&quot;&gt;↑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.45em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.4500000000000002em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-1.8519700000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.45em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:-0.03588em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29011428571428566em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.38327999999999984em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;[&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;ln&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.59803em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.420246em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.75em&quot;&gt;&lt;span style=&quot;top:-1.415892em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em;color:red&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em;color:red&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em;color:red&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:red&quot;&gt;↓&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.75em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-2.102em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.898em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.720216em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:6.646808em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;변분추론의 목적은 &lt;strong&gt;사후분포와 variance distribution의 KL divergence를 줄이는 것&lt;/strong&gt;이었다. 그런데 사후분포가 무엇인지 모르므로 KL divergence를 어떻게 조작해서 줄여야할 지 모른다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;/mstyle&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msub&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_\theta(D)=\textcolor{blue}{ELBO}+\textcolor{red}{D_{KL}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.76666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em;color:blue&quot;&gt;E&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:blue&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em;color:blue&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em;color:blue&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em;color:red&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em;color:red&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;사후분포는 ELBO항과 KL Divergence 항의 합으로 유도할 수 있는데, 이 경우 &lt;strong&gt;&lt;div&gt;KL Divergence를 최소화하기 위해 반대급부로 ELBO 항을 최대화시키는 테크닉이 &lt;code&gt;ELBO법&lt;/code&gt;&lt;/div&gt;&lt;/strong&gt;이다. 이 ELBO항은 구할 수 있으므로(tractable) 이런 방식이 가능하게 된다. 이 기법을 &lt;code&gt;Sandwitch method&lt;/code&gt;라고도 부른다.&lt;/p&gt;&lt;h1 id=&quot;-30&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-30&quot; aria-label=&quot; 30 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이번엔 이 ELBO 항을 나눠보도록 하자.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munder&gt;&lt;munder&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo fence=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;ln&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mstyle mathcolor=&quot;blue&quot;&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mo&gt;↑&lt;/mo&gt;&lt;/mstyle&gt;&lt;/munder&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo&gt;∫&lt;/mo&gt;&lt;mi&gt;ln&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;munder&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mrow&gt;&lt;mi&gt;R&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;u&lt;/mi&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;munder&gt;&lt;munder&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi&gt;F&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
\underbrace{\mathbb{E}_{q_\phi(z|x)}\bigg[\ln\ \frac{p_\theta(x,z)}{q_\theta(z|x)}\bigg]}_{\textcolor{blue}{ELBO\uparrow}}
&amp;amp;= \int \ln\frac{p_\theta(x|z)p(z)}{q_\phi(z|x)}q_\phi(z|x)dz \\
&amp;amp;= \underbrace{\mathbb{E}_{q_\phi (z|x)}[p_\theta(x|z)]}_{Reconstruction \ Term}-\underbrace{D_{KL}(q_\phi(z|x)\Vert p(z))}_{Prior\  Fitting\ Term}
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:7.058793000000001em;vertical-align:-3.2793965000000007em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.7793965000000007em&quot;&gt;&lt;span style=&quot;top:-5.779396500000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.45em&quot;&gt;&lt;span style=&quot;top:-1.165862em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em;color:blue&quot;&gt;E&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:blue&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05017em;color:blue&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em;color:blue&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot; style=&quot;color:blue&quot;&gt;↑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.45em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.4500000000000002em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-1.8519700000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.45em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:-0.03588em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29011428571428566em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.38327999999999984em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;[&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;ln&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.59803em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.420246em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.2191504999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.2793965000000007em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.7793965000000007em&quot;&gt;&lt;span style=&quot;top:-5.779396500000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;margin-right:0.44445em;position:relative;top:-0.0011249999999999316em&quot;&gt;∫&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;ln&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.972108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.2191504999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7499999999999998em&quot;&gt;&lt;span style=&quot;top:-1.2903890000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.00773em&quot;&gt;R&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;u&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mspace mtight&quot;&gt;&lt;span class=&quot;mtight&quot;&gt; &lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7500000000000001em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-1.9687200000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:-0.03588em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29011428571428566em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.38327999999999984em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0312799999999998em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7096109999999998em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.75em&quot;&gt;&lt;span style=&quot;top:-1.3875609999999998em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mspace mtight&quot;&gt;&lt;span class=&quot;mtight&quot;&gt; &lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;F&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace mtight&quot;&gt;&lt;span class=&quot;mtight&quot;&gt; &lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.75em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-2.065892em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.934108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.748547em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.2793965000000007em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Reconstruction Term&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;오토인코더의 reconstruction loss를 최소화한다.&lt;/li&gt;&lt;li&gt;reconstruction loss는 입력값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 인코더를 통해 잠재공간(latent space)으로 이동했다가 다시 디코더로 돌아오는 과정에서 일어나는 loss를 의미한다.&lt;/li&gt;&lt;li&gt;즉, 이상적인 샘플링함수로부터 &lt;strong&gt;얼마나 잘 복원했는지&lt;/strong&gt;를 나타낸다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Prior Fitting Term(Regularization Term)&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;입력값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 잠재공간(latent space)에 점으로 표현했을 때  점들이 이루는 분포-즉 잠재분포(latent distribution)- 로 하여금 사전분포(prior distribution)과 비슷해지도록 강제한다.&lt;/li&gt;&lt;li&gt;이상적인 샘플링함수의 값이 &lt;strong&gt;최대한 사전분포(prior)와 유사한 값을 만들어내도록 condition을 부여&lt;/strong&gt;한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-31&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-31&quot; aria-label=&quot; 31 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;다시 한번 VAE에 대해 되짚어보자.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;목적 : 가지고 있는 이미지 입력값(즉, 관측 데이터) &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 잘 표현할 수 있는 잠재공간(latent space)를 찾는 것&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;그러나 사후분포를 정확히 예측할 수 없으므로, 잠재분포를 사후분포에 근사시키기 위하여 variational inference라는 테크닉을 사용함&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;이를 적용한 오토인코더 모델이 Variational Auto-Encoder(VAE)&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-32&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-32&quot; aria-label=&quot; 32 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;그렇기 때문에 &lt;code&gt;VAE&lt;/code&gt;가 Genarative Model이 될 수 있다. &lt;strong&gt;사전분포와 비슷한 잠재분포를 이루도록 새로운 이미지 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_{new}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 만들기 때문&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;반면, 그냥 &lt;code&gt;AutoEncoder&lt;/code&gt;는 단순히 입력값을 인코딩하여 잠재공간(latent space)로 보냈다가 다시 디코딩하여 어떤 출력값을 얻어내는 모델이므로, &lt;strong&gt;Generative Model이라고 볼 수 없다&lt;/strong&gt;.&lt;/p&gt;&lt;h1 id=&quot;-33&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-33&quot; aria-label=&quot; 33 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;단, VAE에는 &lt;strong&gt;몇 가지 한계점&lt;/strong&gt;이 존재한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;Intractable한 모델이다.&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;VAE는 어떤 입력이 주어졌을 때 likelihood를 측정하여 판별할 수 없다. 즉, explicit한 모델이 아니다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;Prior Fitting Term이 미분가능해야한다.&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;KL Divergence는 그 자체로 적분이 들어가있고, 이 적분이 intractable한 경우가 많다. 따라서 &lt;code&gt;Gaussian prior distribution&lt;/code&gt;을 제외하고는 Closed Form이 잘 나오지 않는다.&lt;ul&gt;&lt;li&gt;Closed Form은 수학적으로 유한한 Term으로 표현할 수 있는 항(또는 식)을 의미한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;그런데 SGD나 Adam같은 Optimizer를 사용하려면 KL Divergence를 포함한 Prior Fitting Term이 미분가능해야한다.&lt;/li&gt;&lt;li&gt;따라서 대부분의 VAE는 Gaussian prior distribution를 사용한다. 그것이 Closed Form이 나오는 몇 안되는 분포이기 때문이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;일반적으로 &lt;code&gt;isotropic Gaussian&lt;/code&gt;을 사용한다.&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;모든 output dimension이 independent한 Gaussian Distribution이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mi mathvariant=&quot;script&quot;&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mfrac&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;/munderover&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msubsup&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msubsup&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;ln&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msubsup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;D_{KL}(q_\phi(z|x)\Vert \mathcal{N}(0,1)) = \frac{1}{2}\sum^D_{i=1}(\sigma^2_{z_i}+\mu^2_{z_i}-\ln(\sigma^2_{z_i})-1)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.14736em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.106005em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.8283360000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.864108em&quot;&gt;&lt;span style=&quot;top:-2.453em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.04398em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.1130000000000004em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34709999999999996em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.211208em;vertical-align:-0.34709999999999996em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.864108em&quot;&gt;&lt;span style=&quot;top:-2.453em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.04398em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.1130000000000004em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34709999999999996em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.211208em;vertical-align:-0.34709999999999996em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;ln&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.864108em&quot;&gt;&lt;span style=&quot;top:-2.453em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3280857142857143em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:-0.04398em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.1130000000000004em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34709999999999996em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;Gaussian prior distribution일 경우 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;D_{KL}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;은 위와 같다. 이 식을 loss function에 집어넣어 학습시키면 원하는 결과가 나온다.&lt;/p&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-34&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-34&quot; aria-label=&quot; 34 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;변분추론의 과정이 잘 이해되지 않는다면, 다음 블로그를 참고하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://ratsgo.github.io/generative%20model/2017/12/19/vi/&quot;&gt;변분추론(Variational Inference)&lt;/a&gt;&lt;/p&gt;&lt;p&gt;AutoEncoder의 모든 것을 아주 잘 설명한 이활석님의 유튜브 강의도 참고해보자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://www.youtube.com/watch?v=o_peo6U7IRM&amp;amp;feature=youtu.be&amp;amp;ab_channel=naverd2&quot;&gt;오토인코더의 모든 것 - 1/3&lt;/a&gt;&lt;/p&gt;&lt;p&gt;해당 영상을 정리한 강의노트도 참고해보자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://deepinsight.tistory.com/127&quot;&gt;[정리노트] [AutoEncoder의 모든것] Chap4. Variational AutoEncoder란 무엇인가(feat. 자세히 알아보자)&lt;/a&gt;&lt;/p&gt;&lt;p&gt;마지막으로, 머리카락으로 예시를 잘 들어놓은 다음 블로그의 글도 참고해보자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;http://jaejunyoo.blogspot.com/2017/04/auto-encoding-variational-bayes-vae-1.html&quot;&gt;초짜 대학원생의 입장에서 이해하는 Auto-Encoding Variational Bayes (VAE) (1)&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-35&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-35&quot; aria-label=&quot; 35 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;adversarial-auto-encoder&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#adversarial-auto-encoder&quot; aria-label=&quot;adversarial auto encoder permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Adversarial Auto-encoder&lt;/h3&gt;&lt;p&gt;앞에서 살펴보았던 Variational Auto-encoder(VAE)의 문제점은, prior fitting term을 미분하기 위해 KL Divergence가 사용된다는 것이었다. 이는 곧 VAE가 사용할 수 있는 사전분포를 Gaussian prior distributtion에 한정시키므로, 다른 분포인 경우에는 활용하기가 힘들었다.&lt;/p&gt;&lt;p&gt;그렇다면 &lt;strong&gt;Gaussian Distribution이 아닌 다른 prior distribution을 활용할 수 없을까&lt;/strong&gt;? 이런 의도에서 나온 것이 &lt;strong&gt;&lt;code&gt;Adversarial Auto-encoder(AAE)&lt;/code&gt;&lt;/strong&gt;다.  사실, AAE는 &lt;strong&gt;VAE의 prior fitting term을 GAN objective로 바꿔버린 것&lt;/strong&gt;에 불과하다.&lt;/p&gt;&lt;h1 id=&quot;-36&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-36&quot; aria-label=&quot; 36 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:305px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:53.90625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;aae&quot; title=&quot;aae&quot; src=&quot;/static/cc2ece6f996fb7ddcfd94c54d03c051e/a3e09/aae.png&quot; srcSet=&quot;/static/cc2ece6f996fb7ddcfd94c54d03c051e/6f3f2/aae.png 256w,/static/cc2ece6f996fb7ddcfd94c54d03c051e/a3e09/aae.png 305w&quot; sizes=&quot;(max-width: 305px) 100vw, 305px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-37&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-37&quot; aria-label=&quot; 37 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;AAE는 잠재분포(latent distribution)를 샘플링한 어떤 분포라도 prior distribution에 맞출 수 있다. &lt;/p&gt;&lt;p&gt;성능도 VAE에 비해서 좋은 경우가 많다.&lt;/p&gt;&lt;h2 id=&quot;gan&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gan&quot; aria-label=&quot;gan permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GAN&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;GAN(Generative Adversarial Network)&lt;/code&gt;&lt;/strong&gt;이라는 방법론은 재미있는 아이디어에서 시작한다.&lt;/p&gt;&lt;h1 id=&quot;-38&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-38&quot; aria-label=&quot; 38 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;위조 지폐를 판별하는 경찰(Discriminator)과, 경찰을 속이려는 위조지폐범(Generator)이 있다고 하자.&lt;/p&gt;&lt;p&gt;경찰은 본인이 가지고 있는 진짜 지폐와 위조 지폐를 비교하여 판독하는데, 이 과정을 거듭할수록 &amp;#x27;진짜&amp;#x27;와 &amp;#x27;가짜&amp;#x27;에 대한 판별을 더욱 더 잘하게 된다.&lt;/p&gt;&lt;p&gt;반면, 위조지폐범은 더욱 더 정교하게 &amp;#x27;진짜&amp;#x27; 같은 &amp;#x27;가짜&amp;#x27;를 만들어 경찰을 속이려 하면서, &amp;#x27;진짜&amp;#x27;와 닮은 지폐를 만들어 낼 수 있게 된다.&lt;/p&gt;&lt;h1 id=&quot;-39&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-39&quot; aria-label=&quot; 39 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;위 과정의 반복을 통해 &lt;strong&gt;궁극적으로 Generator 성능을 더 높이는 것이 GAN의 목적&lt;/strong&gt;이다. 그러나 GAN의 장점은, 무엇보다도 &lt;strong&gt;&lt;div&gt;Generator를 학습시키는 Discriminator가 점차 좋아진다&lt;/div&gt;&lt;/strong&gt;는 데에 있다. &lt;/p&gt;&lt;h1 id=&quot;-40&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-40&quot; aria-label=&quot; 40 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi&gt;min&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi&gt;max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\underset{G}{\min}\ \underset{D}{\max}\ V(D,G) = \mathbb{E}_{x\sim p_{data}(x)}[\log D(x)] + \mathbb{E}_{z\sim p_z(z)}[\log(1- D(G(z)))]&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.494331em;vertical-align:-0.7443310000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.66786em&quot;&gt;&lt;span style=&quot;top:-2.355669em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;min&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.744331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999983em&quot;&gt;&lt;span style=&quot;top:-2.355669em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;max&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7443310000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1052em;vertical-align:-0.3551999999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34480000000000005em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15122857142857138em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3551999999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1052em;vertical-align:-0.3551999999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34480000000000005em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.16454285714285719em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3551999999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;위의 식은 GAN의 알고리즘을 나타내며, 우변은 &lt;code&gt;Loss(생성값과 실제값의 오차)&lt;/code&gt;이다. &lt;strong&gt;&lt;div&gt;Generator는 이 Loss를 최소화하려고 하고, Discriminator는 이 Loss를 최대화하려고 한다.&lt;/div&gt;&lt;/strong&gt;&lt;/p&gt;&lt;h1 id=&quot;-41&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-41&quot; aria-label=&quot; 41 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;gan-vs-vae&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gan-vs-vae&quot; aria-label=&quot;gan vs vae permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GAN vs VAE&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:45.3125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;gan-vs-vae&quot; title=&quot;gan-vs-vae&quot; src=&quot;/static/24cea04b57e6dce40a6c4b3a64aaa77c/2bef9/gan-vs-vae.png&quot; srcSet=&quot;/static/24cea04b57e6dce40a6c4b3a64aaa77c/6f3f2/gan-vs-vae.png 256w,/static/24cea04b57e6dce40a6c4b3a64aaa77c/01e7c/gan-vs-vae.png 512w,/static/24cea04b57e6dce40a6c4b3a64aaa77c/2bef9/gan-vs-vae.png 1024w,/static/24cea04b57e6dce40a6c4b3a64aaa77c/71c1d/gan-vs-vae.png 1536w,/static/24cea04b57e6dce40a6c4b3a64aaa77c/df88b/gan-vs-vae.png 1906w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-42&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-42&quot; aria-label=&quot; 42 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;VAE는 다음과 같은 cycle로 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 생성했다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;학습과정&lt;ul&gt;&lt;li&gt;입력값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 인코더를 통해 latent vector &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 만들고, 이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 다시 디코딩하여 원래 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 복원시키도록 인코더와 디코더의 파라미터를 학습한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;생성과정&lt;ul&gt;&lt;li&gt;잠재분포(latent distribution) &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_\theta(z)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 샘플링한 뒤 디코딩을 하여 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_{new}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 출력한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-43&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-43&quot; aria-label=&quot; 43 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;GAN은 다음과 같은 방식을 취한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;학습과정&lt;ul&gt;&lt;li&gt;z라는 잠재분포(latent distribution)에서 출발하여 Generator를 통해 Fake를 만들어낸다.&lt;/li&gt;&lt;li&gt;Discriminator는 기존의 레이블과 Fake 이미지를 비교하여 판독하며 분류기를 학습한다.&lt;/li&gt;&lt;li&gt;Generator는 Discriminator가 Fake 이미지에 대해 True가 나오도록 위조기를 업데이트하여 학습한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-44&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-44&quot; aria-label=&quot; 44 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;gan의-목표&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gan%EC%9D%98-%EB%AA%A9%ED%91%9C&quot; aria-label=&quot;gan의 목표 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GAN의 목표&lt;/h3&gt;&lt;p&gt;GAN은 결국 generator와 discriminator 간의 &lt;code&gt;minimax 게임&lt;/code&gt;과 같다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;minimax 알고리즘&lt;/code&gt; : 적대적 상황에서 메리트를 최대로, 위기를 최소로 만드는 것을 모티브로 하는 알고리즘&lt;ul&gt;&lt;li&gt;즉, 게임 내에서 두 적대적 모델이 각 턴마다 최적의 결정을 만들어내기 위해 서로 경쟁한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-45&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-45&quot; aria-label=&quot; 45 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;Discriminator의 알고리즘을 수식으로 나타내면 다음과 같다. GAN 모델의 알고리즘에서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 제외하고 만든 식이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi&gt;max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\underset{D}{\max}\ V(G,D) = \mathbb{E}_{x\sim p_{data}}[\log D(x)] + \mathbb{E}_{x\sim p_G}[\log(1- D(x))]&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.494331em;vertical-align:-0.7443310000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999983em&quot;&gt;&lt;span style=&quot;top:-2.355669em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;max&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7443310000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15122857142857138em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3567071428571427em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.14329285714285717em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;Generator가 고정되어있을 때 이 &lt;strong&gt;Disciminator를 최적화(optimalization)시키는 form&lt;/strong&gt;은 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;/msubsup&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;D^*_G(x) = \frac{p_{data}(x)}{p_{data}(x)+p_G(x)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7386959999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mbin mtight&quot;&gt;∗&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.363em;vertical-align:-0.936em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-46&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-46&quot; aria-label=&quot; 46 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;Generator 알고리즘은 다음과 같다. &lt;strong&gt;동일한 Loss에 대해서,  Dicriminator는 최대화하려고 했다면 Generator는 최소화하려고 한다.&lt;/strong&gt;&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi&gt;min&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\underset{G}{\min}\ V(G,D) = \mathbb{E}_{x\sim p_{data}}[\log D(x)] + \mathbb{E}_{x\sim p_G}[\log(1- D(x))]&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.4943309999999999em;vertical-align:-0.744331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.66786em&quot;&gt;&lt;span style=&quot;top:-2.355669em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;min&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.744331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15122857142857138em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3567071428571427em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.14329285714285717em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;h1 id=&quot;-47&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-47&quot; aria-label=&quot; 47 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;Optimal Discriminator를 Generator의 수식에 적용시키면, 다음과 같은 식을 유도할 수 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;/msubsup&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mfrac&gt;&lt;/mfrac&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mfrac&gt;&lt;/mfrac&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mn&gt;4&lt;/mn&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;munder&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo fence=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo fence=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;true&quot;&gt;⏟&lt;/mo&gt;&lt;/munder&gt;&lt;mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mtext&gt;Jenson-Shannon Divergence(JSD)&lt;/mtext&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mn&gt;4&lt;/mn&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;msub&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;J&lt;/mi&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mn&gt;4&lt;/mn&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
V(G,D^*_G(x)) 
&amp;amp;= \mathbb{E}_{x\sim p_{data}}[\log \frac{p_{data}(x)}{p_{data}(x)+p_G(x)}] + \mathbb{E}_{x\sim p_G}[\log\frac{p_G(x)}{p_{data}(x)+p_G(x)}]\\
&amp;amp;= \mathbb{E}_{x\sim p_{data}}[\log \frac{p_{data}(x)}{\frac{p_{data}(x)+p_G(x)}{2}}] + \mathbb{E}_{x\sim p_G}[\log\frac{p_G(x)}{\frac{p_{data}(x)+p_G(x)}{2}}] - \log 4\\
&amp;amp;= \underbrace{D_{KL}\bigg[p_{data},\frac{p_{data}+p_G}{2}\bigg]+D_{KL}\bigg[p_{G},\frac{p_{data}+p_G}{2}\bigg]}_{2\times \textrm{Jenson-Shannon Divergence(JSD)}} - \log4\\
&amp;amp;= 2D_{JSD}[p_{data},p_G]-log4
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:11.38303em;vertical-align:-5.441515000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:5.941514999999999em&quot;&gt;&lt;span style=&quot;top:-7.9645150000000005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7386959999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mbin mtight&quot;&gt;∗&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-5.301514999999999em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.306514999999999em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:1.3315150000000004em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:5.441515000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:5.941514999999999em&quot;&gt;&lt;span style=&quot;top:-7.9645150000000005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15122857142857138em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3567071428571427em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.14329285714285717em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-5.301514999999999em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15122857142857138em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.1099999999999994em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.01em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.01em&quot;&gt;&lt;span style=&quot;top:-2.6550000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.485em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15122857142857138em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3567071428571427em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.14329285714285717em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.345em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2399999999999998em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.01em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.687em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.01em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.245em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3567071428571427em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.14329285714285717em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.1099999999999994em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.01em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.01em&quot;&gt;&lt;span style=&quot;top:-2.6550000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.485em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15122857142857138em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3567071428571427em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.14329285714285717em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.345em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2399999999999998em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.01em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.687em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.01em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.245em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.306514999999999em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.4500000000000002em&quot;&gt;&lt;span style=&quot;top:-1.1269700000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mord text mtight&quot;&gt;&lt;span class=&quot;mord textrm mtight&quot;&gt;Jenson-Shannon Divergence(JSD)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.45em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord munder&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.4500000000000002em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-1.8519700000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;stretchy&quot; style=&quot;height:0.548em;min-width:1.6em&quot;&gt;&lt;span class=&quot;brace-left&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M0 6l6-6h17c12.688 0 19.313.3 20 1 4 4 7.313 8.3 10 13
 35.313 51.3 80.813 93.8 136.5 127.5 55.688 33.7 117.188 55.8 184.5 66.5.688
 0 2 .3 4 1 18.688 2.7 76 4.3 172 5h399450v120H429l-6-1c-124.688-8-235-61.7
-331-161C60.687 138.7 32.312 99.3 7 54L0 41V6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-center&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMidYMin slice&quot;&gt;&lt;path d=&quot;M199572 214
c100.7 8.3 195.3 44 280 108 55.3 42 101.7 93 139 153l9 14c2.7-4 5.7-8.7 9-14
 53.3-86.7 123.7-153 211-199 66.7-36 137.3-56.3 212-62h199568v120H200432c-178.3
 11.7-311.7 78.3-403 201-6 8-9.7 12-11 12-.7.7-6.7 1-18 1s-17.3-.3-18-1c-1.3 0
-5-4-11-12-44.7-59.3-101.3-106.3-170-141s-145.3-54.3-229-60H0V214z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;span class=&quot;brace-right&quot; style=&quot;height:0.548em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;0.548em&quot; viewBox=&quot;0 0 400000 548&quot; preserveAspectRatio=&quot;xMaxYMin slice&quot;&gt;&lt;path d=&quot;M399994 0l6 6v35l-6 11c-56 104-135.3 181.3-238 232-57.3
 28.7-117 45-179 50H-300V214h399897c43.3-7 81-15 113-26 100.7-33 179.7-91 237
-174 2.7-5 6-9 10-13 .7-1 7.3-1 20-1h17z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.45em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;[&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2603300000000002em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;[&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2603300000000002em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.59803em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.49803em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:1.3315150000000004em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.45em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.09618em&quot;&gt;J&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:5.441515000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이 수식을 통해, GAN의 목적은 &lt;strong&gt;실제 데이터의 분포와 학습한 데이터의 분포 사이에 &lt;code&gt;Jenson-Shannon Divergence(JSD)&lt;/code&gt;를 최소화하는 것&lt;/strong&gt;임을 알 수 있다.&lt;/p&gt;&lt;h1 id=&quot;-48&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-48&quot; aria-label=&quot; 48 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;단, 이 수식은 &lt;strong&gt;Discriminator가 Optimal하다는 가정 하에&lt;/strong&gt;, Generator 알고리즘에 대입해야 나오는 식이다. &lt;/p&gt;&lt;p&gt;실제로는 Discriminator가 Optimal Discriminator에 수렴한다는 것을 보장하기힘들고, 이 때 Ganerator가 위의 식과 같이 전개될 수 없다. 따라서 이론적으로는 타당하나 현실적으로는 JSD를 줄이는 방식을 사용하기 힘들기는 하다.&lt;/p&gt;&lt;h1 id=&quot;-49&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-49&quot; aria-label=&quot; 49 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;dcgan&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#dcgan&quot; aria-label=&quot;dcgan permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;DCGAN&lt;/h3&gt;&lt;p&gt;최초에 이안 굿펠로우(Ian Goodfellow)가 발표한 GAN 모델은 MLP모델이었는데, 이를 이미지 도메인으로 개량한 것이 DCGAN이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:321px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:48.828125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;dcgan&quot; title=&quot;dcgan&quot; src=&quot;/static/7c9a9b014d53186c97711f804d65777c/30592/dcgan.png&quot; srcSet=&quot;/static/7c9a9b014d53186c97711f804d65777c/6f3f2/dcgan.png 256w,/static/7c9a9b014d53186c97711f804d65777c/30592/dcgan.png 321w&quot; sizes=&quot;(max-width: 321px) 100vw, 321px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;벡터를 늘리기 위해 Generator에는 Deconvolution을 활용했고, Discriminator에는 Convolution 연산을 수행했다. 알고리즘적으로 개량된 것은 없지만, 하이퍼파라미터나 테크닉에 대한 실험적 결과를 수록하고 있다. &lt;/p&gt;&lt;h1 id=&quot;-50&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-50&quot; aria-label=&quot; 50 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;info-gan&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#info-gan&quot; aria-label=&quot;info gan permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Info-GAN&lt;/h3&gt;&lt;p&gt;Info-GAN은 단순히 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 이용해 이미지만 만들어내는 것이 아니라, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;c&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;c&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;c&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라는 클래스를 집어넣는 방식이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:44.140625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAJCAIAAAC9o5sfAAAACXBIWXMAABYlAAAWJQFJUiTwAAABFklEQVQoz4WQz0rDQBDG83IeFcSDJ0HwCTwpvXvxZgVPBQ+Cb1BECEGkPQQr2EMrAcEGoglJbVOadLOzszMm8Q+yUfxu8+385tsZixtC4qvn/GI0c5+WfBeovk/JqnogMjoto6Ya9hfQGog9Jz+6TideygLUL2wTrjtCwbsub9rZ2i2eB5WjqflFE6buK3fG+XH/5fBenXrcGvL+TdweJG2P/YyMKWbyTHKY6ziTC6Al0MEDbzliw17tuBwX5uIW/6GPhMsJrzuwbYvOGFmj/vdg9EX2Yj4ZwZmnk14gu480jIzoCtZaI2IznIi+N6R6aukoRKUUAJSINZ+nYRRNp2+l9ROTpQAKCaKQqD/jKrikEQGUEOId2mf/vepGDEwAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;info-gan&quot; title=&quot;info-gan&quot; src=&quot;/static/452cf698af3bf9508717efe6bd6e6885/2bef9/info-gan.png&quot; srcSet=&quot;/static/452cf698af3bf9508717efe6bd6e6885/6f3f2/info-gan.png 256w,/static/452cf698af3bf9508717efe6bd6e6885/01e7c/info-gan.png 512w,/static/452cf698af3bf9508717efe6bd6e6885/2bef9/info-gan.png 1024w,/static/452cf698af3bf9508717efe6bd6e6885/71c1d/info-gan.png 1536w,/static/452cf698af3bf9508717efe6bd6e6885/69d6b/info-gan.png 1912w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;C라는 나오는 컨디션벡터(원-핫벡터)를 제공하여 Generator의 multi-model distribution 학습을 돕는다.&lt;/p&gt;&lt;h1 id=&quot;-51&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-51&quot; aria-label=&quot; 51 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;text2image&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#text2image&quot; aria-label=&quot;text2image permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Text2Image&lt;/h3&gt;&lt;p&gt;Text2Image는 문장이 주어지면 이미지를 만드는 모델이다.&lt;/p&gt;&lt;p&gt;문장을 입력으로 받아 Conditional GAN을 통해 이미지를 출력해낸다.&lt;/p&gt;&lt;p&gt;OpenAI에서 제공한 DALL-E의 원조 격이다.&lt;/p&gt;&lt;h1 id=&quot;-52&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-52&quot; aria-label=&quot; 52 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;cyclegan&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#cyclegan&quot; aria-label=&quot;cyclegan permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;CycleGAN&lt;/h3&gt;&lt;p&gt;CycleGAN은 이미지 사이의 도메인을 바꿀 수 있는 모델이다.&lt;/p&gt;&lt;p&gt;어떤 이미지가 주어지면, 그 이미지를 주어진 조건에 알맞은 형태로 변형시킨다. 예를 들어, 그냥 말 사진을 얼룩말 사진으로 바꿔버린다던가, 사진을 고흐의 화풍으로 그려내는 일을 할 수 있다.&lt;/p&gt;&lt;h1 id=&quot;-53&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-53&quot; aria-label=&quot; 53 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이 CycleGAN은 &lt;strong&gt;&lt;code&gt;Cycle-consistency loss&lt;/code&gt;&lt;/strong&gt;라는 아주 중요한 알고리즘을 활용하고 있다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:30.078125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAAAsSAAALEgHS3X78AAAAzklEQVQY05VQSW6EMBDk/w8cRcoFMNhu7xsYQSApjEaaa+rQcndV9eKu1u33n7iuC1FK1X2aW/U6z59931MDiifyBryXZUFx27Zmpq6U4rw/G3LO67oaY7TWMYZ5nqAWQkCjSIHlnIMKISByIbsYo7UWuffeOYv0OA6sUGtV2hCRNgYtbtZaHwIaPYvck51zUGAabFJKdMFwqNH79d2PI6OJJXCp9APrv14kZUr5NtP7ZhxZSn7OPhrgZ7MgpfJ9ZHUhDmzi44Dlnw/Ch/0B7N9XIuSAIJQAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;cycle-consistency-loss&quot; title=&quot;cycle-consistency-loss&quot; src=&quot;/static/2faadd114bfebbc1055358d2592da8f8/2bef9/cycle-consistency-loss.png&quot; srcSet=&quot;/static/2faadd114bfebbc1055358d2592da8f8/6f3f2/cycle-consistency-loss.png 256w,/static/2faadd114bfebbc1055358d2592da8f8/01e7c/cycle-consistency-loss.png 512w,/static/2faadd114bfebbc1055358d2592da8f8/2bef9/cycle-consistency-loss.png 1024w,/static/2faadd114bfebbc1055358d2592da8f8/71c1d/cycle-consistency-loss.png 1536w,/static/2faadd114bfebbc1055358d2592da8f8/0bd45/cycle-consistency-loss.png 1833w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-54&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-54&quot; aria-label=&quot; 54 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;일반적으로는 이미지를 다른 도메인으로 변경하려면 원래 도메인(ex-달리는 말)과 변경하려는 도메인(ex-달리는 얼룩말)의 사진이 모두 필요하다. 그런데 Cycle-consistency loss는 임의의 말 이미지(ex-물 먹는 말, 서 있는 말)들과 임의의 얼룩말 이미지(ex- 물 먹는 얼룩말, 서 있는 얼룩말)들을 잔뜩 모아두고, 이를 이용해 하나의 이미지를 다른 도메인의 이미지로 그냥 변형시켜준다.&lt;/p&gt;&lt;p&gt;이 과정에서 &lt;strong&gt;총 두개의 GAN모델이 사용&lt;/strong&gt;된다.&lt;/p&gt;&lt;h1 id=&quot;-55&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-55&quot; aria-label=&quot; 55 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;star-gan&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#star-gan&quot; aria-label=&quot;star gan permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Star-GAN&lt;/h3&gt;&lt;p&gt;네이버 AI 인턴으로 일하고 있던 대학원생 최윤제님이 개발한 모델이다. &lt;/p&gt;&lt;p&gt;Star-GAN은 단순한 이미지를 원하는 형태로 control할 수 있게 만들어주는 모델이다. 예를 들어 짧은 머리의 20대 남성의 사진을 중년 남성, 또는 여성, 긴 머리의 남성 등으로 만들 수 있다.&lt;/p&gt;&lt;p&gt;후속 논문이 계속 나오고 있을 정도로 실용적인 기술이기도 하다.&lt;/p&gt;&lt;h1 id=&quot;-56&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-56&quot; aria-label=&quot; 56 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;progressive-gan&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#progressive-gan&quot; aria-label=&quot;progressive gan permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Progressive-GAN&lt;/h3&gt;&lt;p&gt;Progressive-GAN은 고해상도의 이미지를 잘 만들 수 있는 GAN이다. &lt;/p&gt;&lt;p&gt;최초에 4x4같은 작은 사이즈부터 시작해 점차 HD size까지 픽셀을 키워나가며 해상도를 점진적으로 상승(progressive)하므로 이런 이름이 붙었다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[딥러닝 베이직 - 07.Transformer 모델 기초]]></title><description><![CDATA[Sequential Models - Transformer by 최성준 교수님, BoostCamp AI Tech 3주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/20_transformer/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/20_transformer/</guid><pubDate>Thu, 04 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;transformer&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#transformer&quot; aria-label=&quot;transformer permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Transformer&lt;/h1&gt;&lt;h2 id=&quot;sequential-model의-한계점&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#sequential-model%EC%9D%98-%ED%95%9C%EA%B3%84%EC%A0%90&quot; aria-label=&quot;sequential model의 한계점 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Sequential Model의 한계점&lt;/h2&gt;&lt;p&gt;&lt;code&gt;RNN&lt;/code&gt;에서 다루었던 Sequential Model들은 완벽한 구성성분을 가진 특정 데이터가 아니면 학습하기가 어려웠다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;문장을 학습시킨다고 할 때&lt;ul&gt;&lt;li&gt;Original Sequence&lt;ul&gt;&lt;li&gt;나는 오늘 학교에 가서 학식을 먹었다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Trimmed Sequence - 문장마다 길이가 다르다.&lt;ul&gt;&lt;li&gt;나는 오늘 학교에 갔다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Ommited Sequence - 문장성분이 누락되어 있을 수 있다.&lt;ul&gt;&lt;li&gt;오늘 학식 먹었다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Permuted Sequence - 성분의 순서가 permute 될 수 있다.&lt;ul&gt;&lt;li&gt;오늘 학교 가서 학식 먹었지, 나는.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이런 문제를 해결하고자 등장한 것이 &lt;strong&gt;&lt;code&gt;Transformer&lt;/code&gt;&lt;/strong&gt; 구조이다.&lt;/p&gt;&lt;h2 id=&quot;transformer-구조&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#transformer-%EA%B5%AC%EC%A1%B0&quot; aria-label=&quot;transformer 구조 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Transformer 구조&lt;/h2&gt;&lt;h3 id=&quot;transformer란&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#transformer%EB%9E%80&quot; aria-label=&quot;transformer란 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Transformer란?&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://arxiv.org/pdf/1706.03762.pdf&quot;&gt;Attention is All you Need, 2017&lt;/a&gt;&lt;/p&gt;&lt;p&gt;재귀적으로 Input이 들어가는 RNN 모델과 달리, Transfomer는 &lt;strong&gt;&lt;code&gt;attention&lt;/code&gt;&lt;/strong&gt;이라고 불리는 구조를 활용하여 sequence를 해석하는 모델이다.&lt;/p&gt;&lt;p&gt;최초에는 &lt;code&gt;신경망기계번역(Neural Machine Translation, NMT)&lt;/code&gt; 문제에 사용되었지만, 시퀀스 데이터를 처리하고 인코딩하는 방법이므로, NMT 문제 뿐 아니라 &lt;strong&gt;여러 분야에 사용될 수 있다&lt;/strong&gt;. 이미지 분류, detection, visual translation 등 활용도가 넓다.&lt;/p&gt;&lt;p&gt;아래의 내용들은 Jay Almmar의 Transformer에 관련된 글을 얕게 풀이한다. 더 깊은 내용이 궁금하다면 해당 블로그를 참조하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;http://jalammar.github.io/illustrated-transformer/&quot;&gt;The Illustrated Transformer&lt;/a&gt;&lt;/p&gt;&lt;h3 id=&quot;형태&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%98%95%ED%83%9C&quot; aria-label=&quot;형태 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;형태&lt;/h3&gt;&lt;p&gt;불어를 입력값으로 주면, 번역한 영어를 내뱉는 &lt;code&gt;seq2seq(Sequence-to-Sequence)&lt;/code&gt; 모델이 있다고 하자.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:65.234375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;en-de-stack&quot; title=&quot;en-de-stack&quot; src=&quot;/static/576e37e4cd6ed2d8923b3e274417e5e2/2bef9/en-de-stack.png&quot; srcSet=&quot;/static/576e37e4cd6ed2d8923b3e274417e5e2/6f3f2/en-de-stack.png 256w,/static/576e37e4cd6ed2d8923b3e274417e5e2/01e7c/en-de-stack.png 512w,/static/576e37e4cd6ed2d8923b3e274417e5e2/2bef9/en-de-stack.png 1024w,/static/576e37e4cd6ed2d8923b3e274417e5e2/c929c/en-de-stack.png 1218w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;기존의 RNN 구조는 단어의 개수만큼 재귀적으로 수행해야만 했다. 그러나, Transformer는 &lt;strong&gt;재귀적 수행이 없고, 한번에 N개의 단어를 모두 처리할 수 있다&lt;/strong&gt;. 물론 Generation 시에는 autoregressive하게 단어마다 수행한다.&lt;/p&gt;&lt;p&gt;또, 동일한 구조를 갖지만, 파라미터가 다르게 학습되는 &lt;strong&gt;&lt;code&gt;인코더(Encoder)&lt;/code&gt;&lt;/strong&gt;와 &lt;strong&gt;&lt;code&gt;디코더(Decoder)&lt;/code&gt;&lt;/strong&gt;가 stack되어있다.&lt;/p&gt;&lt;p&gt;여기서 중요한 포인트는 다음과 같다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;어떻게 N개의 단어를 한번에 처리할 수 있는가?&lt;/li&gt;&lt;li&gt;인코더와 디코더 간에 어떤 정보를 주고받는가?&lt;/li&gt;&lt;li&gt;디코더가 어떻게 Generation 할 수 있는가?&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;특히 그 중 1번에 집중해서 살펴보기로 한다.&lt;/p&gt;&lt;h3 id=&quot;구조-뜯어-보기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B5%AC%EC%A1%B0-%EB%9C%AF%EC%96%B4-%EB%B3%B4%EA%B8%B0&quot; aria-label=&quot;구조 뜯어 보기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;구조 뜯어 보기&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:792px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:51.953125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;encoder-detail&quot; title=&quot;encoder-detail&quot; src=&quot;/static/1da193b987cd1d6838e4665b4c19d548/9a86a/encoder-detail.png&quot; srcSet=&quot;/static/1da193b987cd1d6838e4665b4c19d548/6f3f2/encoder-detail.png 256w,/static/1da193b987cd1d6838e4665b4c19d548/01e7c/encoder-detail.png 512w,/static/1da193b987cd1d6838e4665b4c19d548/9a86a/encoder-detail.png 792w&quot; sizes=&quot;(max-width: 792px) 100vw, 792px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;인코더는 &lt;strong&gt;&lt;code&gt;Self-Attention&lt;/code&gt;&lt;/strong&gt;과 &lt;strong&gt;&lt;code&gt;Feed Forward Neural Network&lt;/code&gt;&lt;/strong&gt;라는 두개의 층을 모아놓은 것으로 이루어져 있다. 인코더는 &lt;strong&gt;모든 단어 벡터를 한번에 입력으로 받으며, 출력값을 바로 다음 인코더로 전달&lt;/strong&gt;한다. 이 중 Feed Forward Nerual Network는 MLP과 별 다른 차이가 없다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:64.0625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;self-attention&quot; title=&quot;self-attention&quot; src=&quot;/static/dbeb1331cff42a9f74fa2ff22148327f/2bef9/self-attention.png&quot; srcSet=&quot;/static/dbeb1331cff42a9f74fa2ff22148327f/6f3f2/self-attention.png 256w,/static/dbeb1331cff42a9f74fa2ff22148327f/01e7c/self-attention.png 512w,/static/dbeb1331cff42a9f74fa2ff22148327f/2bef9/self-attention.png 1024w,/static/dbeb1331cff42a9f74fa2ff22148327f/21335/self-attention.png 1082w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Self-Attention&lt;/code&gt;&lt;/strong&gt;은 &lt;code&gt;Attention 메커니즘&lt;/code&gt;의 핵심이라고 할 수 있다. Self-Attention에서 Feed Forward 층으로 넘어갈 때,  문장에서 특정 단어만 보는 것이 아니라 &lt;strong&gt;&lt;div&gt;문장의 모든 단어를 모두 활용&lt;/div&gt;&lt;/strong&gt;한다. 따라서  Self-Attention 층에서 output을 내보낼 때, 모든 단어들과의 dependancy가 생기게 된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Feed Forward Neural Network는 input된 단어들에 대해 독립적 / 병렬적으로 수행하므로 dependacy가 없다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:437px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:94.53125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;self-attention2&quot; title=&quot;self-attention2&quot; src=&quot;/static/a103df16bceed84e7dd0dac59042db48/5a428/self-attention2.png&quot; srcSet=&quot;/static/a103df16bceed84e7dd0dac59042db48/6f3f2/self-attention2.png 256w,/static/a103df16bceed84e7dd0dac59042db48/5a428/self-attention2.png 437w&quot; sizes=&quot;(max-width: 437px) 100vw, 437px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&amp;quot;The animal didn&amp;#x27;t cross the street because it was too tired.&amp;quot;라는 문장이 있다고 했을 때, it이 무엇을 의미하는 지 알아내는 것은 문장 해석의 중요한 포인트이다.&lt;/p&gt;&lt;p&gt;이 때 Self Attention을 사용하면 &lt;strong&gt;문장 내에서 it에 대응하는 단어들을 모두 고려&lt;/strong&gt;하여 가장 가능성이 높은 &amp;#x27;animal&amp;#x27;로 학습하게 된다.  따라서 기계가 문장 내부에서 단어의 의미를 비교적 더 잘 이해할 수 있다고 표현할 수 있다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:875px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:62.890625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;self-attention3&quot; title=&quot;self-attention3&quot; src=&quot;/static/45dbc2a47b2cd6d2ef8ba28ef2fac164/71c8e/self-attention3.png&quot; srcSet=&quot;/static/45dbc2a47b2cd6d2ef8ba28ef2fac164/6f3f2/self-attention3.png 256w,/static/45dbc2a47b2cd6d2ef8ba28ef2fac164/01e7c/self-attention3.png 512w,/static/45dbc2a47b2cd6d2ef8ba28ef2fac164/71c8e/self-attention3.png 875w&quot; sizes=&quot;(max-width: 875px) 100vw, 875px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;Self Attention 구조는 마치 3개의 신경망처럼 작동한다. 하나의 입력(&lt;strong&gt;&lt;code&gt;단어 임베딩, word embedding&lt;/code&gt;&lt;/strong&gt;)이 주어졌을 때마다 &lt;strong&gt;각 신경망을 통해서 다음과 같은 3개의 벡터를 만들어낸다&lt;/strong&gt;. 이 3개의 벡터를 이용하여 입력받은 단어 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 새로운 벡터로 바꿔주게 된다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Query&lt;/code&gt;&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Key&lt;/code&gt;&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Value&lt;/code&gt;&lt;/strong&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:786px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:94.92187499999999%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;self-attention-output&quot; title=&quot;self-attention-output&quot; src=&quot;/static/087b831f622f83e4529c1bbf646530f0/321ea/self-attention-output.png&quot; srcSet=&quot;/static/087b831f622f83e4529c1bbf646530f0/6f3f2/self-attention-output.png 256w,/static/087b831f622f83e4529c1bbf646530f0/01e7c/self-attention-output.png 512w,/static/087b831f622f83e4529c1bbf646530f0/321ea/self-attention-output.png 786w&quot; sizes=&quot;(max-width: 786px) 100vw, 786px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;단어가 주어지면, 해당 단어에 대한 Query 벡터와 나머지 모든 단어에 대한 Key 벡터를 구한 뒤 내적하여 &lt;strong&gt;&lt;code&gt;Score 벡터&lt;/code&gt;&lt;/strong&gt;를 구한다. 이 score들은 &lt;strong&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째 단어가 다른 단어와 얼마나 유사도(관계)가 있는지 계산하는 역할&lt;/strong&gt;을 한다. 즉, 신경망은 &lt;strong&gt;&lt;div&gt;해당 단어와 나머지 단어 사이에 얼마나 interaction이 일어나야하는지를 학습할 수 있다.&lt;/div&gt;&lt;/strong&gt; 어떤 단어를 더 주의(attention)깊게 볼 지 정하는 것이므로 attention이라는 이름이 붙여졌다.&lt;/p&gt;&lt;p&gt;계산된 Score 벡터에 &lt;code&gt;Normalize&lt;/code&gt;를 한다. 이 때 나누어주는 8이라는 숫자는 Key 벡터의 dimension과 관련이 있다. Key 벡터를 몇 차원(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d_k)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;으로 만드는가는 직접 정해주는 하이퍼파라미터이며, Normalize 시 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msqrt&gt;&lt;msub&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/msqrt&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sqrt{d_k}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.04em;vertical-align:-0.18278000000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord sqrt&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.85722em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;padding-left:0.833em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.81722em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail&quot; style=&quot;min-width:0.853em;height:1.08em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;1.08em&quot; viewBox=&quot;0 0 400000 1080&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M95,702
c-2.7,0,-7.17,-2.7,-13.5,-8c-5.8,-5.3,-9.5,-10,-9.5,-14
c0,-2,0.3,-3.3,1,-4c1.3,-2.7,23.83,-20.7,67.5,-54
c44.2,-33.3,65.8,-50.3,66.5,-51c1.3,-1.3,3,-2,5,-2c4.7,0,8.7,3.3,12,10
s173,378,173,378c0.7,0,35.3,-71,104,-213c68.7,-142,137.5,-285,206.5,-429
c69,-144,104.5,-217.7,106.5,-221
l0 -0
c5.3,-9.3,12,-14,20,-14
H400000v40H845.2724
s-225.272,467,-225.272,467s-235,486,-235,486c-2.7,4.7,-9,7,-19,7
c-6,0,-10,-1,-12,-3s-194,-422,-194,-422s-65,47,-65,47z
M834 80h400000v40h-400000z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.18278000000000005em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 나눠주어 score의 range를 조정해 준다. 이후 Normalize 된 Score에 Softmax를 취해 이를 &lt;strong&gt;&lt;code&gt;attention rate&lt;/code&gt;&lt;/strong&gt;로 바꾸어 준다.&lt;/p&gt;&lt;p&gt;이 &lt;div&gt;&lt;strong&gt;attention rate와 각각의 단어에서 나오는 Value 벡터들의 가중합(weighted sum)&lt;/strong&gt;이 최종적으로 해당 단어의 &lt;strong&gt;&lt;code&gt;인코딩된 벡터(encoded vector)&lt;/code&gt;&lt;/strong&gt;가 된다.&lt;/div&gt;&lt;/p&gt;&lt;p&gt;이 때 유의할 사항들이 몇 가지 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Score 벡터를 구하기 위해 Query 벡터와 Key 벡터를 내적해야 하므로, &lt;strong&gt;Query와 Key 벡터의 차원은 같아야한다.&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;Value 벡터는 차원이 달라도 된다&lt;/strong&gt;. Value 벡터는 attention rate와 가중합을 계산하기만 하면 되기 때문이다.&lt;/li&gt;&lt;li&gt;최종적으로 산출되는 (특정 단어의) &lt;strong&gt;인코딩된 벡터는 Value 벡터의 차원과 동일&lt;/strong&gt;하다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이를 행렬로 표현하면 다음과 같다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:43.359375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAJCAIAAAC9o5sfAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA+0lEQVQoz32QXW7DIBCEff8TtceIKqVvVR8ixdSYsoBh2R9c7KhVorr9JKQFNDAzw/pNaz/jGjnyygVIqnKRpnd3dwz3G4jyCXIxdLIvnt18SSlgmSiCWJDRclz0WBySmpknJxA1LrW1JjU3lZxLLlyppazE7UDcz5aiPkguq3Xw9Pp8TSaN7yU5+3bC5Pdc7T/bN0R7xs1eragih6X8Ke6c/dmTu8wIma5BcHf7u7Shm1HVtnMbRMSiLVzCQkicUEj0uO1ayXxM02Q7ADDPznsffQoQYwjOuQDe2pmZ5ZH+zTCOozEGIDzEFqkbtC2iW1ubfn+hG0SsiPgFDWYOckk1ussAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;self-attention-matrix&quot; title=&quot;self-attention-matrix&quot; src=&quot;/static/954ec4bf60f887427d004fd544c120af/2bef9/self-attention-matrix.png&quot; srcSet=&quot;/static/954ec4bf60f887427d004fd544c120af/6f3f2/self-attention-matrix.png 256w,/static/954ec4bf60f887427d004fd544c120af/01e7c/self-attention-matrix.png 512w,/static/954ec4bf60f887427d004fd544c120af/2bef9/self-attention-matrix.png 1024w,/static/954ec4bf60f887427d004fd544c120af/71c1d/self-attention-matrix.png 1536w,/static/954ec4bf60f887427d004fd544c120af/5b503/self-attention-matrix.png 1976w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 각 row는 하나의 단어를 의미하고, column은 단어의 임베딩을 의미한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Q,K,V&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 각각 Query 벡터, Key 벡터, Value 벡터를 의미한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:893px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:39.0625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAICAYAAAD5nd/tAAAACXBIWXMAAAsSAAALEgHS3X78AAABB0lEQVQoz41S226DMAzt/z/ub6Y9TXvYB3RlF4EmAV2rNrQJSQhJiM8IE3SUTZsjJ74cWz5WVuiFiBYaxekAK8NggzpYsUcYUhSLYDvq/S/sWLMajWuJ8aZ2aISN5Qjeweye4Fs1YYyf0FNsNRYbYyClHF7GKnAu8P54xtvDAd4TXFPjeH8Dyz9QnQS41HjdG+x4O58wXiEE5HmONE2RZRmS5xew6gTJG/BKQusGWknwYg1nFIqixPYocJvU2JTqb8ree1hrJ58xBqX1hWrPgkKYrWdBOS45ns53M+B3MFG48i/5+YS9dj3YqxaHuzVEUv7SkDD/Ffi5IY3TtQ7nTQG9rRZU/iufz9hyTIe+EL4AAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;self-attention-matrix2&quot; title=&quot;self-attention-matrix2&quot; src=&quot;/static/752c1c91e1b4dbca1b64f59a7e026b9b/6c745/self-attention-matrix2.png&quot; srcSet=&quot;/static/752c1c91e1b4dbca1b64f59a7e026b9b/6f3f2/self-attention-matrix2.png 256w,/static/752c1c91e1b4dbca1b64f59a7e026b9b/01e7c/self-attention-matrix2.png 512w,/static/752c1c91e1b4dbca1b64f59a7e026b9b/6c745/self-attention-matrix2.png 893w&quot; sizes=&quot;(max-width: 893px) 100vw, 893px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;길게 설명했던 Self Attention의 과정을 matrix간의 계산으로 표현하면 위와 같은 짧은 식으로 도출된다.&lt;/p&gt;&lt;p&gt;이 때 softmax는 Row-wise하게 적용해야하며, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;dim(V)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;-여기서는 3-와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;Z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;dim(Z)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;Z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 같고, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;dim(Q)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;dim(K)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와는 다를 수 있다는 점을 주목해야한다.&lt;/p&gt;&lt;h3 id=&quot;원리&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9B%90%EB%A6%AC&quot; aria-label=&quot;원리 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;원리&lt;/h3&gt;&lt;p&gt;그렇다면 도대체 왜 이러한 &lt;strong&gt;&lt;code&gt;어텐션 메커니즘(Attention Mechanism)&lt;/code&gt;&lt;/strong&gt;이 잘 작동하는 것일까?&lt;/p&gt;&lt;p&gt;어떤 이미지가 주어졌다고 생각하자. 이 이미지를 MLP나 CNN에 집어넣으면, 입력이 Fix되어 있으므로 출력이 고정된다.&lt;/p&gt;&lt;p&gt;그러나 Transformer는 네트워크가 Fix되어있다고 하더라도, 인코딩하려는 단어와 &lt;strong&gt;그 옆의 단어들에 따라&lt;/strong&gt; 인코딩 결과값이 달라진다. 따라서 결과적으로 MLP보다 좀 더 유연한 모델이 된다. 단순히 &lt;strong&gt;&lt;div&gt;입력된 단어에 따라 출력이 고정되는 것이 아니라, 옆에 주어진 다른 입력(단어)들을 모두 고려하므로 출력이 변화할 여지가 있기 때문&lt;/div&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;물론, 더 많은 것을 고려하므로 더 많은 연산(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n^2)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.064108em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 필요하게 된다. 예를 들어 1000개의 단어가 주어지면 한번에 1000^2의 입력을 처리해야한다. 반면, RNN은 1000개의 단어가 주어지면 그냥 연산을 1000번 수행하면 된다. &lt;strong&gt;너무 긴 입력이 주어지면 메모리 부족으로 한번에 계산할 수 없다는 한계를 가진다.&lt;/strong&gt; &lt;/p&gt;&lt;p&gt;정리하자면,&lt;/p&gt;&lt;ol&gt;&lt;li&gt;기존의 RNN에 비해 훨씬 더 많은 문맥정보들을 파악할 수 있다.&lt;/li&gt;&lt;li&gt;연산자원이 비교적 많이(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n^2)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.064108em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 든다.&lt;/li&gt;&lt;/ol&gt;&lt;h3 id=&quot;multi-headed-attention&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#multi-headed-attention&quot; aria-label=&quot;multi headed attention permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Multi-headed attention&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Multi-headed attention(MHA)&lt;/code&gt;&lt;/strong&gt;은 앞에서 이야기했던 attention 과정을 여러번 수행한다. &lt;strong&gt;하나의 임베딩 벡터(입력)에 대해서 (Query,Key,Value) 벡터 셋을 여러 개 만든다&lt;/strong&gt;. 이 때문에 multi-headed(머리가 여러개)라는 이름이 붙었다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:55.46875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;mha&quot; title=&quot;mha&quot; src=&quot;/static/9a721b7e3b77140f0a51e6cb38117209/2bef9/mha.png&quot; srcSet=&quot;/static/9a721b7e3b77140f0a51e6cb38117209/6f3f2/mha.png 256w,/static/9a721b7e3b77140f0a51e6cb38117209/01e7c/mha.png 512w,/static/9a721b7e3b77140f0a51e6cb38117209/2bef9/mha.png 1024w,/static/9a721b7e3b77140f0a51e6cb38117209/b79a5/mha.png 1372w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;MHA로 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 개의 헤드를 사용해 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 개의 인코딩된 벡터가 나오게 되면, 이 값을 concatenate하여 다음 인코더로 넘겨준다. 이 때, 여러 번의 인코더를 타야 하므로, 인코딩된 벡터(출력값)의 차원을 인코딩 하기 전의 입력값 차원과 동일하게 맞춰줄 필요가 있다. 그런데 MHA의 (concatenate 된) 출력 벡터는 기존의 입력값에 비해 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 배의 차원이 되었으므로, 이를 적절한 행렬 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^O&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8413309999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8413309999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와의 행렬곱을 통해 기존의 크기로 변환시켜준다.&lt;/p&gt;&lt;p&gt;그러나 &lt;strong&gt;실제 구현은 위와 같은 방식으로 이루어지지 않는다&lt;/strong&gt;. 원래 주어진 임베딩 단어의 차원이 100차원이고 10개의 head를 사용한다면, 이를 10개로 나눈다. 즉, 10차원짜리 입력만을 가지고 Query, Key, Value 벡터를 만들게 된다. 이는 구현된 코드를 보고 직접 확인해보자.&lt;/p&gt;&lt;h3 id=&quot;positional-encoding&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#positional-encoding&quot; aria-label=&quot;positional encoding permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Positional Encoding&lt;/h3&gt;&lt;p&gt;Transformer 모델은 마치 bias처럼 입력에 특정 값을 더해주는 &lt;strong&gt;&lt;code&gt;Positional Encoding&lt;/code&gt;&lt;/strong&gt;도 수행한다. 왜 이런 Positional Encoding이 필요한 것일까?&lt;/p&gt;&lt;p&gt;Transformer 모델은 여러 단어들을 동시에 고려하도록 설계되었지만, &lt;strong&gt;sequential한 정보가 포함되어있지 않기 때문&lt;/strong&gt;이다. [a,b,c,d]로 이루어진 문장구조나, [a,d,b,c]로 이루어진 문장 구조나 동일하게 단어간의 attention만을 측정할 뿐이므로, &lt;strong&gt;순서에 독립적(order-independent)&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;그러나 실제 문장에는 어떤 단어가 먼저 나왔고 그렇지 않은지가 중요하다. 이미지에서도 마찬가지다. 따라서 주어진 입력에 대해 어떤 값을 더함으로써 이러한 문제를 해결한다.&lt;/p&gt;&lt;p&gt;이에 대해서는 자세히 설명된 다음 블로그 글을 참고하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://skyjwoo.tistory.com/entry/positional-encoding%EC%9D%B4%EB%9E%80-%EB%AC%B4%EC%97%87%EC%9D%B8%EA%B0%80&quot;&gt;positional encoding이란 무엇인가&lt;/a&gt;&lt;/p&gt;&lt;p&gt;위에서 이야기했던 Transformer 모델의 Encoder 과정들을 다음과 같은 도식으로 표현할 수 있다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:56.640625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;transformer-layer&quot; title=&quot;transformer-layer&quot; src=&quot;/static/6d335c4bd879cc8ccbd1ee6da1b90846/2bef9/transformer-layer.png&quot; srcSet=&quot;/static/6d335c4bd879cc8ccbd1ee6da1b90846/6f3f2/transformer-layer.png 256w,/static/6d335c4bd879cc8ccbd1ee6da1b90846/01e7c/transformer-layer.png 512w,/static/6d335c4bd879cc8ccbd1ee6da1b90846/2bef9/transformer-layer.png 1024w,/static/6d335c4bd879cc8ccbd1ee6da1b90846/d9ed0/transformer-layer.png 1415w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;decoder&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#decoder&quot; aria-label=&quot;decoder permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Decoder&lt;/h2&gt;&lt;p&gt;이제  Decoder를 살펴보자.&lt;/p&gt;&lt;p&gt;인코더에서 디코더로 넘어갈 때, 어떤 정보들이 전달될까?&lt;/p&gt;&lt;h3 id=&quot;원리-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9B%90%EB%A6%AC-1&quot; aria-label=&quot;원리 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;원리&lt;/h3&gt;&lt;p&gt;&lt;img src=&quot;http://jalammar.github.io/images/t/transformer_decoding_1.gif&quot; alt=&quot;decoding_1&quot;/&gt;&lt;/p&gt;&lt;p&gt;Input에 있는 값들을 디코더에 있는, 즉 출력하고자 하는 단어들에 매핑하는 &lt;code&gt;Attention Map&lt;/code&gt;을 만드려면,  Key 벡터와 Value 벡터가 필요하다. 마지막 인코더 층(가장 상위 인코더)의 출력값은 여러 단어들의 [Key 벡터, Value 벡터] 시리즈로 제공되고, 이 값들을 각 Decoder 내부의 &amp;#x27;&lt;strong&gt;&lt;code&gt;Encoder-Decoder Attention&lt;/code&gt;&lt;/strong&gt;&amp;#x27; 계층에서 해석하여 입력 시퀀스의 적절한 위치를 잡아주게 된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 Encoder-Decoder Attention Layer는 MHA와 동일한 방식으로 동작하지만, &lt;strong&gt;Query 벡터를 이전 Decoder에서 받아오고 Key, Value 벡터를 인코더 스택에서 받아온다&lt;/strong&gt;는 차이가 있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;img src=&quot;http://jalammar.github.io/images/t/transformer_decoding_2.gif&quot; alt=&quot;decoding_2&quot;/&gt;&lt;/p&gt;&lt;p&gt;디코더의 마지막 층에서는, &lt;strong&gt;디코더 스택의 출력물을 단어들의 분포로 만들어 내보낸다&lt;/strong&gt;. 이 단어들을 매번 샘플링하는 형식으로 모델의 출력이 만들어진다.&lt;/p&gt;&lt;p&gt;모든 출력이 끝나면, &lt;code&gt;EOS(End Of Sentence) 토큰&lt;/code&gt;이 나와 종료를 알린다.&lt;/p&gt;&lt;h2 id=&quot;transformer의-활용&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#transformer%EC%9D%98-%ED%99%9C%EC%9A%A9&quot; aria-label=&quot;transformer의 활용 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Transformer의 활용&lt;/h2&gt;&lt;p&gt;최근에는 Transformer의 활용이 시퀀스 데이터에만 머무르지 않고, 이미지 영역 등 다양한 분야에 사용되고 있다.&lt;/p&gt;&lt;h3 id=&quot;vision-transformer&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#vision-transformer&quot; aria-label=&quot;vision transformer permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Vision Transformer&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://arxiv.org/pdf/2010.11929.pdf&quot;&gt;An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale&lt;/a&gt;&lt;/p&gt;&lt;p&gt;Transformer 모델에서 인코더를 차용해 이미지 분류 시에 사용한다.&lt;/p&gt;&lt;h3 id=&quot;dall-e&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#dall-e&quot; aria-label=&quot;dall e permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;DALL-E&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://openai.com/blog/dall-e/&quot;&gt;DALL·E: Creating Images from Text&lt;/a&gt;&lt;/p&gt;&lt;p&gt;문장이 주어지면 이에 맞는 이미지를 만들어낸다.  Transformer의 decoder만 이용하여 만들었다.(GPT-3를 활용했다고 한다)&lt;/p&gt;&lt;hr/&gt;&lt;p&gt;개인적으로 3주간 공부하면서 가장 어려운 파트였다. 개념 상 이해해야 하는 부분도 많았고, 특히 최근에 가장 많이 사용되는 모델링이어서 제대로 이해하고싶은 마음에 이것저것 찾아보는데 무엇하나 명확하게 그림이 그려지지 않아서 더 어려웠다. 아쉬운대로 넘어가지만, 추후에 좀 더 자세히 살펴보며 보강해가야겠다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[RNN의 원리]]></title><description><![CDATA[RNN 첫걸음 by 임성빈 교수님, BoostCamp AI Tech 3주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/18_rnn_math/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/18_rnn_math/</guid><pubDate>Thu, 04 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;recurrent-neural-network&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#recurrent-neural-network&quot; aria-label=&quot;recurrent neural network permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Recurrent Neural Network&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Recurrent Neural Network(RNN, 순환신경망)&lt;/code&gt;&lt;/strong&gt;는 이전까지 다뤘던 CNN 등과는 달리, &lt;strong&gt;시계열 데이터(Time-Series Data) 또는 시퀀스 데이터(Sequence Data)를 다루는 데에 주로 사용되는 모델&lt;/strong&gt;이다.&lt;/p&gt;&lt;h2 id=&quot;시퀀스-데이터&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%8B%9C%ED%80%80%EC%8A%A4-%EB%8D%B0%EC%9D%B4%ED%84%B0&quot; aria-label=&quot;시퀀스 데이터 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;시퀀스 데이터&lt;/h2&gt;&lt;h3 id=&quot;시퀀스-데이터란&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%8B%9C%ED%80%80%EC%8A%A4-%EB%8D%B0%EC%9D%B4%ED%84%B0%EB%9E%80&quot; aria-label=&quot;시퀀스 데이터란 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;시퀀스 데이터란&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;시퀀스 데이터&lt;/code&gt;&lt;/strong&gt;란 &lt;strong&gt;순차적으로 들어오는 데이터&lt;/strong&gt;를 일컫는다. 소리, 문자열, 주가 등이 있다.&lt;/p&gt;&lt;p&gt;시퀀스 데이터에서는 정보들이 시간 순서대로 종합되어서 시점별로 나타난다. 따라서 &lt;strong&gt;이벤트의 발생 순서가 데이터의 중요한 요소로서 동작&lt;/strong&gt;한다.&lt;/p&gt;&lt;p&gt;기존의 데이터에 비해 시퀀스 데이터에서 유의할 점은, &lt;strong&gt;&lt;div&gt;독립동등분포(independant identically distributed, i.i.d)를 위배하기 쉽다&lt;/div&gt;&lt;/strong&gt;는 것이다. 맥락과 순서가 중요하기 때문에, 순서를 바꾸거나 과거 정보에 손실이 발생하면, 데이터의 확률 분포가 바뀌거나 미래 예측의 정확도가 떨어지게 된다.&lt;/p&gt;&lt;h3 id=&quot;시퀀스-데이터를-다루는-방법&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%8B%9C%ED%80%80%EC%8A%A4-%EB%8D%B0%EC%9D%B4%ED%84%B0%EB%A5%BC-%EB%8B%A4%EB%A3%A8%EB%8A%94-%EB%B0%A9%EB%B2%95&quot; aria-label=&quot;시퀀스 데이터를 다루는 방법 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;시퀀스 데이터를 다루는 방법&lt;/h3&gt;&lt;p&gt;이전 시퀀스의 정보를 가지고 앞으로 발생할 데이터의 확률분포를 다루기 위하여, &lt;code&gt;조건부확률&lt;/code&gt;을 이용할 수 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;msub&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;/msub&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/munderover&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
P(X1_,\dots,X_t)&amp;amp;=P(X_t|X_1,\dots,X_{t-1})P(X_1,\dots,X_{t-1})\\
&amp;amp;= P(X_t|X_1,\dots,X_{t-2})P(X_1,\dots,X_{t-2})\times P(X_1,\dots,X_{t-2})\\
&amp;amp;=\prod^t_{s=1}P(X_s|X_{s-1},\dots,X_1)
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:6.3476740000000005em;vertical-align:-2.9238370000000007em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.423837em&quot;&gt;&lt;span style=&quot;top:-6.364398em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.7805610000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:-0.07610800000000001em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.864398em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.7805610000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.423837em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.7805610000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.9238370000000007em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.423837em&quot;&gt;&lt;span style=&quot;top:-6.364398em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.7805610000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.864398em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.7805610000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.423837em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.7805610000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7805610000000003em&quot;&gt;&lt;span style=&quot;top:-1.882887em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∏&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.267113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.9238370000000007em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;현재 시점인 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;  까지의 확률분포를 알기 위해, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t-1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69841em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;까지의 확률분포를 구한다. &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t-1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69841em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 확률분포를 구하기 위해 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t-2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69841em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;까지의 확률분포를 구한다. 이를 반복하여 최초까지 내려가면, 최초의 시점부터 순차적으로 조건부확률을 곱하는 식으로 일반화할 수 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_t \sim P(X_t|X_{t-1},\dots,X_1)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;다만, 위의 조건부확률과는 달리 시퀀스 데이터를 분석할 때 일반적으로는 과거의 &lt;strong&gt;모든 정보(즉, 최초시점부터의 모든 정보)를 사용하지는 않는다&lt;/strong&gt;. 최근 어느 시점까지의 정보를 이용하거나, 몇 개의 과거 정보들을 truncation하기도 한다. 필요성에 따라서 어느 시점까지 데이터를 활용할 지 모델링이 달라질 수 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_{t+1} \sim P(X_{t+1}| \textcolor{red}{X_t,X_{t-1}\dots,X_1})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.891661em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em;color:red&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:red&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em;color:red&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:red&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot; style=&quot;color:red&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:red&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em;color:red&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;따라서, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_{t+1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.891661em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 예측하기 위한 &lt;strong&gt;조건부확률에 들어가는 데이터 길이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_{t},X_{t-1},\dots,X_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.891661em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 가변적&lt;/strong&gt;이므로, 시퀀스 데이터를 다루기 위해서는 길이가 가변적인 데이터를 다룰 수 있는 모델이 필요하다.&lt;/p&gt;&lt;p&gt;가변적인 길이나 아니라 고정된 길이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;τ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tau&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.1132em&quot;&gt;τ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;만큼의 시퀀스만 사용하는 경우 &lt;strong&gt;&lt;code&gt;AR 모델&lt;/code&gt;&lt;/strong&gt;, 즉 &lt;strong&gt;&lt;code&gt;자기회귀모델(Autoregressive Model)&lt;/code&gt;&lt;/strong&gt;이라고 부른다. 다만, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;τ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tau&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.1132em&quot;&gt;τ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 길이는 직접 정해주어야하는 하이퍼파라미터가 되므로, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;τ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tau&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.1132em&quot;&gt;τ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 결정할 때에는 어느정도의 설계자의 사전지식이 필요하게 된다.&lt;/p&gt;&lt;p&gt;또, 문제에 따라 먼 과거의 정보들을 고려해야 하기도하고, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;τ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tau&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.1132em&quot;&gt;τ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 바뀌는 경우가 있다. 이런 경우 사용되는 방법이 RNN의 기본 모형인 &lt;strong&gt;&lt;code&gt;잠재 AR 모델(Latent Autoregressive Models)&lt;/code&gt;&lt;/strong&gt;, 즉 &lt;strong&gt;&lt;code&gt;잠재자기회귀모델&lt;/code&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mo&gt;→&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mo&gt;→&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;_t \sim P(X_t|X_{t-1},\textcolor{red}{\dots,X_1}) \textcolor{red}{\rarr H_t}\\
X_{t+1} \sim P(X_{t+1}|X_t,\textcolor{red}{X_{t-1}\dots,X_1})\textcolor{red}{\rarr H_{t+1}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.5168699999999999em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot; style=&quot;color:red&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:red&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em;color:red&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot; style=&quot;color:red&quot;&gt;→&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em;color:red&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.891661em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em;color:red&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:red&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot; style=&quot;color:red&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot; style=&quot;color:red&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em;color:red&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot; style=&quot;color:red&quot;&gt;→&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.891661em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em;color:red&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;color:red&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot; style=&quot;color:red&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이 경우 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 를 예측하기 위해 직전정보인 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t-1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69841em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 시점의 정보 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_{t-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.891661em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 제외한 나머지 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_{t-2},\dots&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.891661em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 정보들을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;H_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라는 &lt;strong&gt;&lt;code&gt;잠재변수&lt;/code&gt;&lt;/strong&gt;로 인코딩하여 이용한다.&lt;/p&gt;&lt;p&gt;따라서, 직전 시점의 정보와 잠재변수&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;H&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 두 가지만 가지고 예측할 수 있기 때문에 고정된 길이의 데이터를 가지고 모델링 할 수 있게 된다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mtext&gt;Net&lt;/mtext&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;H_t = \textrm{Net}_\theta(H_{t-1}, X_{t-1})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;Net&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이 때 잠재변수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;H_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 신경망을 통해 반복 사용하여 시퀀스 데이터의 패턴을 학습하는 모델을 &lt;strong&gt;&lt;code&gt;RNN&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;/p&gt;&lt;h2 id=&quot;rnn-이해하기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#rnn-%EC%9D%B4%ED%95%B4%ED%95%98%EA%B8%B0&quot; aria-label=&quot;rnn 이해하기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;RNN 이해하기&lt;/h2&gt;&lt;h3 id=&quot;rnn의-순전파&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#rnn%EC%9D%98-%EC%88%9C%EC%A0%84%ED%8C%8C&quot; aria-label=&quot;rnn의 순전파 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;RNN의 순전파&lt;/h3&gt;&lt;p&gt;가장 기본적인 RNN 모형은 MLP와 유사한 모형이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
O_t &amp;amp;= H_tW^{(2)} + b^{(2)}\\
H_t &amp;amp;= \sigma(X_tW^{(1)}+b^{(1)})
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.196em;vertical-align:-1.348em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.848em&quot;&gt;&lt;span style=&quot;top:-3.91em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.3120000000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.348em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.848em&quot;&gt;&lt;span style=&quot;top:-3.91em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.3120000000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.348em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이 때, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^{(1)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8879999999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;과 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^{(2)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8879999999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 시퀀스와 상관없이 불변인 가중치 행렬이다.&lt;/p&gt;&lt;p&gt;이 모델의 문제점은, 입력이 현재 시점(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)과 관련된 것만 들어오기 때문에, &lt;strong&gt;예측할 때 과거 정보를 다룰 수 없다&lt;/strong&gt;는 것이다. 따라서 과거 정보를 다루기 위해 RNN에서는 과거의 잠재변수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;H_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 복제하여 다음 순서의 잠재변수를 인코딩하는 데에 사용한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;msubsup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/msub&gt;&lt;msubsup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
O_t &amp;amp;= H_tW^{(2)} + b^{(2)}\\
H_t &amp;amp;= \sigma(X_tW_X^{(1)}+H_{t_1}W_H^{(1)}+b^{(1)})
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.3028000000000004em;vertical-align:-1.4014000000000002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.9014000000000002em&quot;&gt;&lt;span style=&quot;top:-4.0082em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0448em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.3034em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0448em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.4014000000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.9014000000000002em&quot;&gt;&lt;span style=&quot;top:-4.0082em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0448em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.3034em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0448em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4064690000000004em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.293531em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31731428571428577em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2501em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4064690000000004em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.293531em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.4014000000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:33.203125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA5ElEQVQY021R2W6DMBD0/39e2+dKhSoQbAIGH7s+M+BIRW1HljzeY/awyDmHEJjZWhtOgNf/AHuMsXEiSikJ51zff0sp394/hmG83yccBBljIOS9R1BLmB/LqrfGx3ECF4HDbRiJWCn11fXIkVItywLFddVKzcRMjHYCp8Ix7cbGlDmWkLKAjCf+vE16N9u2q/mh9bbvELGonzIS6iShPGufV8Nd12/GWZ84nsmAI/aYKfygTe48h3i0XWo1jko5gku7ahXXx18cbRM1PpzTXb0CbmzLWne14gvQRnzhtbBfNbDIJ5b5luIPfDJCAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;rnn_model&quot; title=&quot;rnn_model&quot; src=&quot;/static/15a44b23e935c294f3892920d80ea3c0/2bef9/rnn_model.png&quot; srcSet=&quot;/static/15a44b23e935c294f3892920d80ea3c0/6f3f2/rnn_model.png 256w,/static/15a44b23e935c294f3892920d80ea3c0/01e7c/rnn_model.png 512w,/static/15a44b23e935c294f3892920d80ea3c0/2bef9/rnn_model.png 1024w,/static/15a44b23e935c294f3892920d80ea3c0/71c1d/rnn_model.png 1536w,/static/15a44b23e935c294f3892920d80ea3c0/a878e/rnn_model.png 2048w,/static/15a44b23e935c294f3892920d80ea3c0/f0293/rnn_model.png 2250w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;유의할 점은, &lt;strong&gt;가중치 행렬 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 세 개 나온다&lt;/strong&gt;는 것이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_X^{(1)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.338331em;vertical-align:-0.29353099999999993em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4064690000000004em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29353099999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 입력 데이터에서 선형 모델을 통해 잠재변수로 인코딩하는 가중치 행렬&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_H^{(1)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.338331em;vertical-align:-0.29353099999999993em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4064690000000004em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29353099999999993em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 이전 시점으로 잠재변수로부터 정보를 받아서 현재시점의 잠재변수로 인코딩하는 가중치 행렬&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^{(2)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8879999999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 만든 잠재변수를 출력값으로 만들어주는 가중치 행렬&lt;/li&gt;&lt;/ul&gt;&lt;div&gt;&lt;p&gt;세 가중치 행렬은 시점 t에 따라 변하지 않음을 명심하자. 위의 세 가중치 행렬들은 동일하게 각각의 시점에서 활용되어 모델링에 사용된다.&lt;/p&gt;&lt;/div&gt;&lt;h3 id=&quot;rnn의-역전파&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#rnn%EC%9D%98-%EC%97%AD%EC%A0%84%ED%8C%8C&quot; aria-label=&quot;rnn의 역전파 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;RNN의 역전파&lt;/h3&gt;&lt;p&gt;RNN의 역전파는 &lt;strong&gt;잠재변수의 연결그래프에 따라 순차적으로 계산&lt;/strong&gt;한다. 이를 &lt;strong&gt;&lt;code&gt;Backpropagation Through Time(BPTT)&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:30.46875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA1ElEQVQY02WQDYvDMAiG+///4d2xg2Pr2qX5MGqSJmnac2SDwh5EXkRf0eE4DmMsAFwuv+N494jTPO/7LvVzFpBoWXTXSikAP4i6jXdrrdbm++tHPRatrTSNNyk6pRbi0AdyWpleOiLFlIfu/TBgiR0ZxxqC9eyemkzdakrZAU7TDEo2+L/r6AG9smsqQ3fikDQCZRcSYfBxDRyJE+ecI3LbWi4lcazEpW6l1KM1idfwJ/1URHTGvu9cEP25ZwghysO8xyZmJ4pQ6zNkT7c7PU/YtvYP0tJayoZrlEIAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;rnn-back&quot; title=&quot;rnn-back&quot; src=&quot;/static/294e47a6a16f47fab6d5041345495df1/2bef9/rnn-back.png&quot; srcSet=&quot;/static/294e47a6a16f47fab6d5041345495df1/6f3f2/rnn-back.png 256w,/static/294e47a6a16f47fab6d5041345495df1/01e7c/rnn-back.png 512w,/static/294e47a6a16f47fab6d5041345495df1/2bef9/rnn-back.png 1024w,/static/294e47a6a16f47fab6d5041345495df1/71c1d/rnn-back.png 1536w,/static/294e47a6a16f47fab6d5041345495df1/a878e/rnn-back.png 2048w,/static/294e47a6a16f47fab6d5041345495df1/0c1ff/rnn-back.png 2192w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;RNN의 모든 시점에서의 예측이 전부 이루어진 다음, 마지막 시점의 Gradient가 타고 올라와서 과거까지 흐른다. &lt;strong&gt;&lt;div&gt;RNN은 모든 시간 스텝에서 파라미터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 공유하므로, 한 시점에서 오류가 역전파되면 이것이 모든 이전시점으로 시간을 거슬러 퍼지게 된다.&lt;/div&gt;&lt;/strong&gt; 그래서 BPTT라는 이름이 붙었다.&lt;/p&gt;&lt;p&gt;위의 그림에서 빨간 선이 역전파과정인데, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;H_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 역전파에 들어오는 Gradient는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;H_{t+1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.891661em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.08125em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;과 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서 각각 하나씩 총 두개가 있다. &lt;/p&gt;&lt;h3 id=&quot;bptt-이해하기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#bptt-%EC%9D%B4%ED%95%B4%ED%95%98%EA%B8%B0&quot; aria-label=&quot;bptt 이해하기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;BPTT 이해하기&lt;/h3&gt;&lt;p&gt;BPTT를 통해 RNN의 가중치행렬 미분을 계산해보면, 아래와 같은 미분의 곱으로 이루어진 항이 계산된다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:36.71875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA3UlEQVQY031QyXLFIAzL///ju3SmM2QBswYbAqnIm9L2Uh9AgIUkL/d9t9au7wJOKdVaAXvv91OZxTrvffQhGmOBiZxIWcZb5pROVEyJmdd1c87hiC/uf2uQoVAfzdY6CES2lApVfAQXaJimhqHr+kMGIcahCQs5Z9iCrIhg3Q79uR7Guv3QmmjbD+f8D3kGm2Ae+Tw5JqUphJBjDEQIzVKGw6d5md3QgTIsYAPA+vF67auCGQxDmL0hqzWGN56ZBxmzUUppbd5h+lMApdYUI6RyOhtGgLQi7VdgXHwB4ECYGAPuc8UAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;bptt&quot; title=&quot;bptt&quot; src=&quot;/static/23b0f8a8e74e3dc766ec0148ced5fd83/2bef9/bptt.png&quot; srcSet=&quot;/static/23b0f8a8e74e3dc766ec0148ced5fd83/6f3f2/bptt.png 256w,/static/23b0f8a8e74e3dc766ec0148ced5fd83/01e7c/bptt.png 512w,/static/23b0f8a8e74e3dc766ec0148ced5fd83/2bef9/bptt.png 1024w,/static/23b0f8a8e74e3dc766ec0148ced5fd83/71c1d/bptt.png 1536w,/static/23b0f8a8e74e3dc766ec0148ced5fd83/a878e/bptt.png 2048w,/static/23b0f8a8e74e3dc766ec0148ced5fd83/691b3/bptt.png 2156w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;최종적으로 나오게 되는 Product Term(빨간색 박스가 나타내는 항)은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i+1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.74285em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;부터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 시점까지의 모든 잠재변수에 대한 미분값이 곱해지는데, 이 때 &lt;strong&gt;&lt;div&gt;시퀀스의 길이(즉, 시점간의 거리)가 길어질수록 해당 항이 불안정해지기 쉬워진다.&lt;/div&gt;&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;매 시점마다 1보다 큰 수가 곱해진다면 미분값이 아주 커진다.&lt;/li&gt;&lt;li&gt;매 시점마다 1보다 작은 수가 곱해진다면 미분값이 아주 작아진다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;가장 주의해야 하는 것은 작은 수의 미분값이 곱해져 &lt;strong&gt;Gradient가 0으로 소실되는 현상(&lt;code&gt;Vanishing Gradient&lt;/code&gt;)&lt;/strong&gt;이다. 이 경우 과거시점으로 갈 수록 Gradient가 점점 작아지기 때문에 과거 시점의 정보가 제대로 반영이 되지 않아 유실된다. 이러면 긴 시퀀스를 분석해야하거나 문맥적인 부분이 중요한 모델의 경우 좋은 결과를 얻을 수 없다.&lt;/p&gt;&lt;p&gt;따라서 시퀀스 길이가 너무 길어지는 경우 길이를 끊는 것이 필요한데, 이를 &lt;strong&gt;&lt;code&gt;truncated BPTT&lt;/code&gt;&lt;/strong&gt;라고 부른다. 과거의 일정 시점마다 블럭을 나눠서 해당 블럭 단위로만 backpropagation 연산을 하는 방법이다.&lt;/p&gt;&lt;p&gt;그러나 이 방법이 완전히 기울기 소실 문제를 해결하지 못하므로, 일반적으로는 Vanilla RNN을 사용하지 않고 좀 더 발전한 RNN( &lt;code&gt;LSTM&lt;/code&gt;, &lt;code&gt;GRU&lt;/code&gt;)을 사용한다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[딥러닝 베이직 - 06. RNN의 기초와 역사 훑어보기]]></title><description><![CDATA[Sequential Models - RNN by 최성준 교수님, BoostCamp AI Tech 3주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/19_rnn_basic/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/19_rnn_basic/</guid><pubDate>Thu, 04 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;rnn의-기초와-변천사&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#rnn%EC%9D%98-%EA%B8%B0%EC%B4%88%EC%99%80-%EB%B3%80%EC%B2%9C%EC%82%AC&quot; aria-label=&quot;rnn의 기초와 변천사 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;RNN의 기초와 변천사&lt;/h1&gt;&lt;p&gt;&lt;code&gt;MLP&lt;/code&gt;는 벡터를 다른 벡터로 바꾸는 것이었고, &lt;code&gt;CNN&lt;/code&gt;은 이미지를 원하는 형태로 바꿔주는것이었다면, &lt;strong&gt;&lt;code&gt;RNN&lt;/code&gt;&lt;/strong&gt;은 시퀀셜 모델을 다루는 것이다.&lt;/p&gt;&lt;h2 id=&quot;sequential-model&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#sequential-model&quot; aria-label=&quot;sequential model permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Sequential Model&lt;/h2&gt;&lt;p&gt;시퀀스 데이터를 처리하는데에 가장 어려운 것은, &lt;strong&gt;&lt;div&gt;길이가 언제 끝날 지 모른다는 것&lt;/div&gt;&lt;/strong&gt;이다. 따라서 &lt;strong&gt;받아들여야하는 입력의 차원을 알 수가 없다&lt;/strong&gt;. 시간이 지날수록, 고려해야하는 과거의 정보량이 늘어난다.&lt;/p&gt;&lt;h3 id=&quot;autoregressive-model&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#autoregressive-model&quot; aria-label=&quot;autoregressive model permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Autoregressive Model&lt;/h3&gt;&lt;p&gt;이를 가장 간단히 해결하는 방법은, &lt;strong&gt;고정된 길이(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;τ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\tau)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.1132em&quot;&gt;τ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 과거 정보만을 확인&lt;/strong&gt;하는 &lt;code&gt;Markov Model&lt;/code&gt;이다. 이를 극단적으로 간단히 만든 것이 바로 직전 시점 정보만을 고려하는 &lt;code&gt;AR(1) 모델&lt;/code&gt;이다.&lt;/p&gt;&lt;h3 id=&quot;latent-autoregressive-model&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#latent-autoregressive-model&quot; aria-label=&quot;latent autoregressive model permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Latent Autoregressive Model&lt;/h3&gt;&lt;p&gt;기존 AR 모델이 직전정보까지밖에 고려하지 못했기 때문에, 이를 보완하여 그 이전 과거의 정보들을 &amp;#x27;기억&amp;#x27;할 수 있는 새로운 AR 모델이 나오게 되었다. 이를 &lt;strong&gt;&lt;code&gt;Latent Autoregressive Model&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:40.625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAICAIAAAB2/0i6AAAACXBIWXMAABYlAAAWJQFJUiTwAAAA60lEQVQY021Ri26DMBDj//9xmrqOSl1JLuRBLq/LTFirbcJC4Bwyto+pD5RSw7bFyMaszjlMckrrulrrxttCRNZaEem/MB0PTFtrtVZjDL6CCY5EZhu8SwcPIfS/+BGXXDQRnJXW0DMzfBalYA7uvb8/HhaJpJ87p5TgrDV5H5Aj56KUhqyJYL4sCvzcGSF5QGsNt71zzui/d64VegJ37n9neaINMCfIcIwxLl938/6G6tVQJI2rbaE61zgeYad+BqyXc/afF56vuVRh7ocn7tL6s/o0z7fLx1UpBduXGBw/juOeA1s8NcCOvgFpw9OcMNJOjAAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;latent-ar&quot; title=&quot;latent-ar&quot; src=&quot;/static/f13be5ceb45c17782e001c4f9ab7b1c1/2bef9/latent-ar.png&quot; srcSet=&quot;/static/f13be5ceb45c17782e001c4f9ab7b1c1/6f3f2/latent-ar.png 256w,/static/f13be5ceb45c17782e001c4f9ab7b1c1/01e7c/latent-ar.png 512w,/static/f13be5ceb45c17782e001c4f9ab7b1c1/2bef9/latent-ar.png 1024w,/static/f13be5ceb45c17782e001c4f9ab7b1c1/71c1d/latent-ar.png 1536w,/static/f13be5ceb45c17782e001c4f9ab7b1c1/c6720/latent-ar.png 1922w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;이 모델의 포인트는 hidden state(또는 latent state) &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 있다. &lt;strong&gt;출력값(다음 시점의 정보)은 입력값(해당 시점의 정보)에 그 전까지의 모든 시점정보들을 요약(summary)한 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 고려&lt;/strong&gt;하여 만들어진다. &lt;/p&gt;&lt;h3 id=&quot;recurrent-neural-network&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#recurrent-neural-network&quot; aria-label=&quot;recurrent neural network permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Recurrent Neural Network&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:26.171875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;rnn-unrolled&quot; title=&quot;rnn-unrolled&quot; src=&quot;/static/34a870b0e60d513e7153b3f27fa66786/2bef9/rnn-unrolled.png&quot; srcSet=&quot;/static/34a870b0e60d513e7153b3f27fa66786/6f3f2/rnn-unrolled.png 256w,/static/34a870b0e60d513e7153b3f27fa66786/01e7c/rnn-unrolled.png 512w,/static/34a870b0e60d513e7153b3f27fa66786/2bef9/rnn-unrolled.png 1024w,/static/34a870b0e60d513e7153b3f27fa66786/71c1d/rnn-unrolled.png 1536w,/static/34a870b0e60d513e7153b3f27fa66786/a878e/rnn-unrolled.png 2048w,/static/34a870b0e60d513e7153b3f27fa66786/8d294/rnn-unrolled.png 2706w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;RNN을 시간순으로 풀면(un-roll) 위와 같은 모형도가 나오게 된다. RNN처럼 Recurrent(되풀이) 구조가 있는 모델을 시간순으로 풀게 되면, 결국 (과거의 입력들이 같이 들어오므로) &lt;strong&gt;입력이 굉장히 많은 네트워크&lt;/strong&gt;로 볼 수 있게 된다.&lt;/p&gt;&lt;p&gt;문제는 과거의 정보들을 미래의 정보로 끌고오기 때문에, 역설적으로 더 오래된(멀리있는) 정보일수록 살아남기가 힘들다는 것이다. 마치 메멘토에서 주인공이 가까운 과거밖에 기억하지 못하는것처럼, 근시간의 정보가 아니면 고려하지 못하게 된다. &lt;strong&gt;&lt;div&gt;RNN은 이처럼 &lt;code&gt;Short-term dependencies&lt;/code&gt;는 잘 잡을 수 있지만, &lt;code&gt;Long-term dependencies&lt;/code&gt;는 잘 잡지 못한다&lt;/div&gt;&lt;/strong&gt;는 치명적인 단점이 있다.&lt;/p&gt;&lt;p&gt;그렇다면 RNN 학습이 도대체 왜 어려운 것일까?&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;U&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;U&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;U&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle mathcolor=&quot;red&quot;&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;/mstyle&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;U&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;U&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;U&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;U&lt;/mi&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/msup&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mo lspace=&quot;0em&quot; rspace=&quot;0em&quot;&gt;⋯&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
h_1 &amp;amp;= \textcolor{red}{\phi(W^Th_0}+U^Tx_1)\\
h_2 &amp;amp;= \textcolor{red}{\phi(W^T\phi(W^Th_0}+U^Tx_1)+U^Tx_2)\\
h_3 &amp;amp;= \textcolor{red}{\phi(W^T\phi(W^T\phi(W^T\phi(W^T}h_0+U^Tx_1)+U^Tx_2)+U^Tx_1)+U^Tx_3)\\
\cdots
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:6.153993em;vertical-align:-2.8269964999999995em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.3269965em&quot;&gt;&lt;span style=&quot;top:-5.4356655em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.8843345000000005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.333003500000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-0.8330035000000007em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.8269964999999995em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.3269965em&quot;&gt;&lt;span style=&quot;top:-5.4356655em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;ϕ&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;U&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.8843345000000005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;ϕ&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;ϕ&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;color:red&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;U&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;U&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.333003500000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;ϕ&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;ϕ&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;ϕ&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;color:red&quot;&gt;ϕ&lt;/span&gt;&lt;span class=&quot;mopen&quot; style=&quot;color:red&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot; style=&quot;color:red&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em;color:red&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;U&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;U&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;U&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;U&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3269964999999992em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;RNN은 이런식으로 과거의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 들을 고려하는 중첩된 구조이다. &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\phi&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 활성화함수(activation function)이다.  시퀀스 길이가 늘어남에 따라, 이처럼 &lt;strong&gt;중첩되는 가중치와 activation function이 굉장히 많아진다&lt;/strong&gt;. &lt;/p&gt;&lt;p&gt;위의 식에서 만약 활성화함수가 &lt;code&gt;sigmoid&lt;/code&gt;라고 하자. sigmoid의 성질은 값을 계속 0과 1사이로 바꿈으로써 축소시키는 것이므로, 함수가 중첩될수록 점점 &lt;strong&gt;&lt;code&gt;vanishing gradient&lt;/code&gt;&lt;/strong&gt;의 문제가 생긴다.&lt;/p&gt;&lt;p&gt;만약, sigmoid가 아닌 &lt;code&gt;ReLU&lt;/code&gt;함수라면 어떨까? ReLU함수는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;&amp;gt;&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&amp;gt;0&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.5782em;vertical-align:-0.0391em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;&amp;gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 일때 해당 input을 그대로 가져가므로, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 input의 곱이 계속 쌓이는 구조가 될 것이다. 따라서 자칫하면 Gradient가 아주 커져 네트워크가 터지는 &lt;strong&gt;&lt;code&gt;exploding gradient&lt;/code&gt;&lt;/strong&gt;의 문제가 생긴다.&lt;/p&gt;&lt;h3 id=&quot;lstm&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#lstm&quot; aria-label=&quot;lstm permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;LSTM&lt;/h3&gt;&lt;p&gt;이러한 Vanilla RNN 단점을 해결, 즉 &lt;code&gt;Long-term dependencies&lt;/code&gt;를 확보하기 위해 만들어진 모델이 &lt;strong&gt;&lt;code&gt;Long Short term Memory(LSTM)&lt;/code&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:44.921875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;lstm&quot; title=&quot;lstm&quot; src=&quot;/static/f0b9c261ec3751de0bbd6b8a68ff2d15/2bef9/lstm.png&quot; srcSet=&quot;/static/f0b9c261ec3751de0bbd6b8a68ff2d15/6f3f2/lstm.png 256w,/static/f0b9c261ec3751de0bbd6b8a68ff2d15/01e7c/lstm.png 512w,/static/f0b9c261ec3751de0bbd6b8a68ff2d15/2bef9/lstm.png 1024w,/static/f0b9c261ec3751de0bbd6b8a68ff2d15/71c1d/lstm.png 1536w,/static/f0b9c261ec3751de0bbd6b8a68ff2d15/99e81/lstm.png 1946w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위는 RNN과 대비되는 LSTM의 모형인데, 언뜻 보면 아주 복잡해보인다.&lt;/p&gt;&lt;p&gt;LSTM의 핵심 아이디어는 &lt;strong&gt;&lt;code&gt;Cell State&lt;/code&gt;&lt;/strong&gt;이다. 컨베이어 벨트로 이해하면 쉬운데, 매 시점마다 컨베이어벨트로 과거 시점의 정보들이 죽 전달되고, 각 시점에서 [해당 입력값을 넣을 것인지 말 것인지], [어떤 정보를 summary에 추가할 것인지], [출력값으로 얼마만큼 내보낼 것인지]를 &lt;code&gt;Gate&lt;/code&gt;에서 결정한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:42.1875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAICAIAAAB2/0i6AAAACXBIWXMAABYlAAAWJQFJUiTwAAABCElEQVQY03VRS1PCMBDm//8PD9566cEDyAFHRaoO7QjS0R5QhFLbJDXvbOqWjo4I7GQymW/3eyTpNafLe4/7crF+mmSM66Lkn5VgX+p3oHeEA+CEBCGV5qZRH9v3PN8gDgAoBzvF/2RWy4oKUqttQZNxfz7tPy6G8fp6lITTlztCFGGSMukPyYhYB8aCMQ6gid9uRmkwmJyPn8NweJauYg8NtnBgj9ydlbYVEaiNq6x4NBvczoOrJHjILqLlZfIaUXSmbTQu9JHYqALdlby/n6U5IdII7bQBZazGOG23Nd5z9tZarbVzrpXYoUm22hS1s9CJnvqIlkwpLTEr53/b+Lb2p/yBQAd+A2xLy8IH/p41AAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;lstm_detail&quot; title=&quot;lstm_detail&quot; src=&quot;/static/2245e2ca41ec45d53986fe938ff148b5/2bef9/lstm_detail.png&quot; srcSet=&quot;/static/2245e2ca41ec45d53986fe938ff148b5/6f3f2/lstm_detail.png 256w,/static/2245e2ca41ec45d53986fe938ff148b5/01e7c/lstm_detail.png 512w,/static/2245e2ca41ec45d53986fe938ff148b5/2bef9/lstm_detail.png 1024w,/static/2245e2ca41ec45d53986fe938ff148b5/71c1d/lstm_detail.png 1536w,/static/2245e2ca41ec45d53986fe938ff148b5/a878e/lstm_detail.png 2048w,/static/2245e2ca41ec45d53986fe938ff148b5/5e1f2/lstm_detail.png 2134w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 시퀀스 데이터로 만든 현재 시점의 입력값 벡터&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 출력값(이자 hidden state)&lt;/li&gt;&lt;li&gt;&lt;code&gt;Previous cell state&lt;/code&gt; : 출력값으로 나가지는 않고, 매 시점마다 과거 시점들의 입력정보들을 linear하게 취합/전달하여 보여주는 값. &lt;code&gt;Forget Gate&lt;/code&gt;에 의해 제어된다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;Previous hidden state&lt;/code&gt; : 이전 시점의 출력값.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 때, Gate의 존재에 주목하자. 총 3개의 게이트가 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;Forget Gate&lt;/code&gt; : 얼마만큼 지울(버릴) 것인지&lt;ul&gt;&lt;li&gt;해당 정보를 버릴 것인지, 아니면 살려서 전달할 것인지 결정한다. 현재 입력 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 이전 출력 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_{t-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.902771em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 입력으로 받아 sigmoid를 적용시키므로 0과 1 사이의 값을 갖게된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;Input Gate&lt;/code&gt;: 무엇을 올릴 것인지&lt;ul&gt;&lt;li&gt;해당 정보 중 어느 것을 cell state에 저장(추가)할 것인지 정한다. &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_{t-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.902771em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 입력으로 받아 sigmoid를 적용시킨 값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.80952em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 곱해서 정보를 취사선택한다. tanh를 적용시킨(출력값은 -1과 1 사이) 이번 시점의 Cell state &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 만들어 지금까지의 Cell state에 섞어서 업데이트한다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;Update Cell&lt;/code&gt; : 직전까지의 정보를 Summary한 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_{t-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.891661em;vertical-align:-0.208331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.301108em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.208331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 Forget Gate를 통과한 값을 곱하고,  이번 시점의 Input Gate를 통과한 값을 더하여 새로운 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;C_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 업데이트한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;Output Gate&lt;/code&gt; : 얼마만큼 내보낼 것인지&lt;ul&gt;&lt;li&gt;Update한 cell state를 한번 더 조작하여 어떤 값을 밖으로 내보낼 지 결정한다. Output Gate만큼 곱해서(element-wise multiplication) 현재의 아웃풋 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2805559999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 만들어낸다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 과정을 풀이해놓은 블로그를 참조하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;http://colah.github.io/posts/2015-08-Understanding-LSTMs/&quot;&gt;Understanding LSTM Networks&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://dgkim5360.tistory.com/entry/understanding-long-short-term-memory-lstm-kr&quot;&gt;Long Short-Term Memory (LSTM) 이해하기&lt;/a&gt;&lt;/p&gt;&lt;h3 id=&quot;gru&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gru&quot; aria-label=&quot;gru permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GRU&lt;/h3&gt;&lt;p&gt;기존의 LSTM이 너무 복잡한 구조를 가지고 있어, 이를 조금 더 단순하게 만든 모델로 뉴욕대 조경현 교수가 제안한 알고리즘이다. &lt;strong&gt;&lt;code&gt;Gated Recurrent Unit&lt;/code&gt;&lt;/strong&gt;의 약자로, &lt;strong&gt;게이트가 3개 있던 LSTM과는 달리 2개의 게이트(reset, update)만을 가진다. 또, cell state가 없고, hidden state만 가진다.&lt;/strong&gt;&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:27.734375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA8ElEQVQY001QXWvCMBT1//+O7WX4MoXii65zWD86pagoIpugSNVWVtaahCa52Wnith7CvSc39+R+NEwNmhTBkDaG4vgkpURQKQWiavjLbzgXn4/jVW95GC12wfwjWO6GQehfzgkR5XmepmmWZUVRgMMiH/FKTKZyrMz9qBVue5vTZDDzXqPWPtnipZSlEOLb4mbBOf8XGysuiffnXqf/MF40w3X77f3xmG4Q55xBcLUAQWVn8WNdLLzBU3f67Eft6ac/278kxQFxJCGbMYaCKI4rOIib/D4zlnL9ukgSOEIyIW9KV9vSFmjSrUr/wrX9AzGVUghD/0GoAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;gru&quot; title=&quot;gru&quot; src=&quot;/static/75d989c345d4d1e164c72c146c42f2f4/2bef9/gru.png&quot; srcSet=&quot;/static/75d989c345d4d1e164c72c146c42f2f4/6f3f2/gru.png 256w,/static/75d989c345d4d1e164c72c146c42f2f4/01e7c/gru.png 512w,/static/75d989c345d4d1e164c72c146c42f2f4/2bef9/gru.png 1024w,/static/75d989c345d4d1e164c72c146c42f2f4/71c1d/gru.png 1536w,/static/75d989c345d4d1e164c72c146c42f2f4/d2782/gru.png 1864w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;code&gt;Reset Gate&lt;/code&gt;가 기존의 &lt;code&gt;Forget Gate&lt;/code&gt;역할을 하고, &lt;code&gt;Input&amp;amp;Output Gate&lt;/code&gt;가 합쳐져 &lt;code&gt;Update Gate&lt;/code&gt; 역할을 한다고 볼 수 있다.&lt;/p&gt;&lt;p&gt;파라미터 개수가 LSTM보다 적음에도 불구하고 비슷한 작용을 하므로, &lt;strong&gt;대체로 일반화 성능이 좋은 편&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;그러나 최근에는 LSTM과 GRU 모두 Transformer가 나오면서 대체되고 있는 추세다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[CNN의 원리]]></title><description><![CDATA[CNN 첫걸음 by 임성빈 교수님, BoostCamp AI Tech 3주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/15_cnn_math/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/15_cnn_math/</guid><pubDate>Wed, 03 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;cnn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#cnn&quot; aria-label=&quot;cnn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;CNN&lt;/h1&gt;&lt;p&gt;기존의 &lt;code&gt;다층신경망(MLP)&lt;/code&gt;는 각 뉴런들이 선형모델과 활성함수로 &lt;strong&gt;모두 연결된(fully connected)&lt;/strong&gt; 구조였다.&lt;/p&gt;&lt;p&gt;이런 다층신경망의 경우는 &lt;strong&gt;각 성분 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_i = \sigma(\sum^p_{j=1}W_{ij}x_j)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.24011em;vertical-align:-0.43581800000000004em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.804292em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2029000000000005em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43581800000000004em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대응하는 가중치 행 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 필요&lt;/strong&gt;했다.&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;convolution-연산-이해하기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#convolution-%EC%97%B0%EC%82%B0-%EC%9D%B4%ED%95%B4%ED%95%98%EA%B8%B0&quot; aria-label=&quot;convolution 연산 이해하기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Convolution 연산 이해하기&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Convolution 연산&lt;/code&gt;&lt;/strong&gt;은 기존의 MLP 방식과는 달리, &lt;strong&gt;&lt;code&gt;커널(kernel, 필터)&lt;/code&gt;&lt;/strong&gt;을 &lt;strong&gt;입력벡터 상에서 움직여가면서&lt;/strong&gt; 선형 모델과 합성함수가 적용되는 구조이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/munderover&gt;&lt;msub&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_i = \sigma\Bigg(\sum^k_{j=1}V_jx_{i+j-1}\Bigg)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.2498900000000006em;vertical-align:-1.4137769999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.8361130000000006em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.4137769999999998em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.22222em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;h_i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째 히든레이어를 의미한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 커널(필터) 사이즈를 의미한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;V&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 가중치 행렬, 즉 커널을 의미한다. &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;V&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;V&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.22222em&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 모든 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;j&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.85396em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해서 같다.&lt;ul&gt;&lt;li&gt;이것이 기존 방식과의 가장 큰 차이점이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;수학적-의미&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%88%98%ED%95%99%EC%A0%81-%EC%9D%98%EB%AF%B8&quot; aria-label=&quot;수학적 의미 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;수학적 의미&lt;/h3&gt;&lt;p&gt;Convolution 연산의 수학적 의미는 &lt;strong&gt;커널을 이용해 신호(signal)을 국소적으로 증폭 혹은 감소시켜 정보를 추출 또는 필터링&lt;/strong&gt; 하는 것이다. 엄밀히 말하면 CNN에서 사용하는 연산은 convolution이 아니고 &lt;strong&gt;&lt;code&gt;cross-correlation&lt;/code&gt;&lt;/strong&gt;이라고 부르지만, 관습적으로 Convolution으로 부르곤 한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mtext&gt;continuous&lt;/mtext&gt;&lt;mspace width=&quot;1em&quot;&gt;&lt;/mspace&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mo&gt;∫&lt;/mo&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;R&lt;/mi&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;d&lt;/mi&gt;&lt;/msup&gt;&lt;/msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mo&gt;∫&lt;/mo&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;R&lt;/mi&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;d&lt;/mi&gt;&lt;/msup&gt;&lt;/msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mtext&gt;discrete&lt;/mtext&gt;&lt;mspace width=&quot;1em&quot;&gt;&lt;/mspace&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;Z&lt;/mi&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;d&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;Z&lt;/mi&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;d&lt;/mi&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
\textrm{continuous} \quad [f*g](x) &amp;amp;= \int_\mathbb{R^d}f(z)g(x+z)dz = \int_\mathbb{R^d}f(x+z)g(z)dz = [g*f](x)\\
\textrm{discrete} \quad [f*g](i) &amp;amp;= \sum_{a\in \mathbb{Z^d}}f(a)g(i+a) = \sum_{a\in\mathbb{Z^d}}f(i+a)g(a) = [g*f](i)
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:5.31275em;vertical-align:-2.4063749999999997em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.9063750000000006em&quot;&gt;&lt;span style=&quot;top:-4.906375000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3600000000000003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;continuous&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:1em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∗&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.6444200000000007em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3600000000000003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;discrete&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:1em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∗&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.4063749999999997em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.9063750000000006em&quot;&gt;&lt;span style=&quot;top:-4.906375000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3600000000000003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;margin-right:0.44445em;position:relative;top:-0.0011249999999999316em&quot;&gt;∫&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:-0.36453em&quot;&gt;&lt;span style=&quot;top:-1.7880500000000001em;margin-left:-0.44445em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathbb mtight&quot;&gt;R&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7820285714285713em&quot;&gt;&lt;span style=&quot;top:-2.786em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9119499999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;margin-right:0.44445em;position:relative;top:-0.0011249999999999316em&quot;&gt;∫&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:-0.36453em&quot;&gt;&lt;span style=&quot;top:-1.7880500000000001em;margin-left:-0.44445em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathbb mtight&quot;&gt;R&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7820285714285713em&quot;&gt;&lt;span style=&quot;top:-2.786em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9119499999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∗&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.6444200000000007em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3600000000000003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.7865749999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathbb mtight&quot;&gt;Z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7820285714285713em&quot;&gt;&lt;span style=&quot;top:-2.786em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.390795em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.7865749999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathbb mtight&quot;&gt;Z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7820285714285713em&quot;&gt;&lt;span style=&quot;top:-2.786em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.390795em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∗&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.4063749999999997em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x+z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.66666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 또는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i+a&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.74285em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 들어간 함수값은 신호를 의미한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 또는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;a&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;만 들어간 함수 값은 커널을 의미한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;연속변수이든 이산변수이든, 합과 적분의 차이를 제외하고는 연산이 거의 동일하다.&lt;/p&gt;&lt;p&gt;CNN을 수식만으로 이해하기는 어려우므로, graphical하게 이해해보자.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:35.546875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAABEUlEQVQY011PW07DMBDMITgZf5wGiUtxBX54lNCQd0JRoDRx7NjxY+3EZRtUKWW0sla7M7Pj4LiCtVaI0WjQWo8SKGWC3Q70sWme9vvmcGgNANL8mR+sxWbB0s4Gjs5SmK6Uueds4pwZA+i+5l+IAUBKifZaGy6MVlzq61E9WIOhTsNBSA0OQznn/ovneVZKIQmfUWo5aslnKcaObNquJaTf/XRVWX8mWU8ZgL0Qo5+1J0vvPcY3ShiZOTv15IsNFGMj/HLDLwjAOvwLFjZ4jWEypSgd6ACc79ruhtA70n53BO8z3K35wSbcPr+8hm9RWX8UVZ3lZVZUSZrHSbaNkvc4jdMwy4sImwQ3NdL+Ki+qX9b+i3h1U+t0AAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;convolution_graph&quot; title=&quot;convolution_graph&quot; src=&quot;/static/bed0438111426e0ed285b9a3454bf92e/2bef9/convolution_graph.png&quot; srcSet=&quot;/static/bed0438111426e0ed285b9a3454bf92e/6f3f2/convolution_graph.png 256w,/static/bed0438111426e0ed285b9a3454bf92e/01e7c/convolution_graph.png 512w,/static/bed0438111426e0ed285b9a3454bf92e/2bef9/convolution_graph.png 1024w,/static/bed0438111426e0ed285b9a3454bf92e/71c1d/convolution_graph.png 1536w,/static/bed0438111426e0ed285b9a3454bf92e/a878e/convolution_graph.png 2048w,/static/bed0438111426e0ed285b9a3454bf92e/b5c21/convolution_graph.png 2154w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;커널은 &lt;strong&gt;정의역 내에서 움직여도 변하지 않고(translation invariant), 주어진 신호에 국소적(local)으로 적용&lt;/strong&gt;된다. 이를 &lt;code&gt;locality&lt;/code&gt;가 있다고 말한다.&lt;/p&gt;&lt;h3 id=&quot;영상처리&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%98%81%EC%83%81%EC%B2%98%EB%A6%AC&quot; aria-label=&quot;영상처리 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;영상처리&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://webnautes.tistory.com/1044&quot;&gt;영상처리 강좌 3 - 컨볼루션(Convolution)과 스무딩(Smoothing), 샤프닝(Sharpening)&lt;/a&gt;&lt;/p&gt;&lt;p&gt;위 링크에서와 같이, Convolution Network의 커널을 적용할 때 어느 위치를 찍느냐에 따라 하나의 이미지를 여러 형태(blur, emboss, outline 등)로 해석할 수 있다.&lt;/p&gt;&lt;h3 id=&quot;다차원에서의-convolution-연산&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%8B%A4%EC%B0%A8%EC%9B%90%EC%97%90%EC%84%9C%EC%9D%98-convolution-%EC%97%B0%EC%82%B0&quot; aria-label=&quot;다차원에서의 convolution 연산 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;다차원에서의 Convolution 연산&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;div&gt;데이터의 성격에 따라 사용하는 커널을 다르게&lt;/div&gt;&lt;/strong&gt; 하며, 다차원으로 Convolution 연산을 할 수 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mtext&gt;1D-conv&lt;/mtext&gt;&lt;mspace width=&quot;1em&quot;&gt;&lt;/mspace&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/munderover&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mtext&gt;2D-conv&lt;/mtext&gt;&lt;mspace width=&quot;1em&quot;&gt;&lt;/mspace&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mtext&gt;3D-conv&lt;/mtext&gt;&lt;mspace width=&quot;1em&quot;&gt;&lt;/mspace&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
\textrm {1D-conv} \quad [f*q](i) &amp;amp;= \sum^d_{p=1}f(p)g(i+p)\\
\textrm {2D-conv} \quad [f*q](i,j) &amp;amp;= \sum_{p,q}f(p,q)g(i+p,j+q) \\
\textrm {3D-conv} \quad [f*q](i,j,k) &amp;amp;= \sum_{p,q,r}f(p,q,r)g(i+p,j+q,k+r)
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:9.01157em;vertical-align:-4.255785em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:4.755785em&quot;&gt;&lt;span style=&quot;top:-6.755785em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.836113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;1D-conv&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:1em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∗&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.002559em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.836113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;2D-conv&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:1em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∗&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.2664409999999997em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.836113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;3D-conv&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:1em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∗&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:4.255785em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:4.755785em&quot;&gt;&lt;span style=&quot;top:-6.755785em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.836113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.836113em&quot;&gt;&lt;span style=&quot;top:-1.8828869999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.300005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.403221em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.002559em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.836113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000003em&quot;&gt;&lt;span style=&quot;top:-1.8999949999999999em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.386113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.2664409999999997em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.836113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000003em&quot;&gt;&lt;span style=&quot;top:-1.8999949999999999em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.386113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:4.255785em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;여기서&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 커널, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;g&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 입력을 뜻한다.&lt;/li&gt;&lt;li&gt;데이터의 종류&lt;ul&gt;&lt;li&gt;1차원 데이터 - 음성, 텍스트 등&lt;/li&gt;&lt;li&gt;2차원 데이터 - 흑백 영상 등&lt;/li&gt;&lt;li&gt;3차원 데이터 - 컬러 영상 등&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;다만, 중요한 것은 &lt;strong&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i,j,k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 위치가 바뀌더라도 커널 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 값은 바뀌지 않는다&lt;/strong&gt;는 점이다.&lt;/p&gt;&lt;h2 id=&quot;2차원-convolution-연산-이해하기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#2%EC%B0%A8%EC%9B%90-convolution-%EC%97%B0%EC%82%B0-%EC%9D%B4%ED%95%B4%ED%95%98%EA%B8%B0&quot; aria-label=&quot;2차원 convolution 연산 이해하기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;2차원 Convolution 연산 이해하기&lt;/h2&gt;&lt;p&gt;2차원에서는 1차원과 달리 입력데이터의 형태가 &lt;strong&gt;행렬형태&lt;/strong&gt;이다. 따라서 1차원에서처럼 한쪽 방향으로 한칸씩 움직이면서 커널을 이동시켰던것과 달리, &lt;strong&gt;커널을 x축과 y축 방향으로 한칸씩 움직여가며 적용&lt;/strong&gt;시킨다.&lt;/p&gt;&lt;p&gt;입력 크기를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(H,W)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, 커널 크기를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(K_H,K_W)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, 출력 크기를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(O_H,O_W)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라 하면, 출력 크기는 다음과 같이 계산한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
O_H &amp;amp;= H-K_H+1\\
O_W &amp;amp;= W-K_W+1
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.0000000000000004em;vertical-align:-1.2500000000000002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7500000000000002em&quot;&gt;&lt;span style=&quot;top:-3.91em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.41em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2500000000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7500000000000002em&quot;&gt;&lt;span style=&quot;top:-3.91em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.41em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07153em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2500000000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;가령, 28x28 입력을 3x3 커널로 2차원 Convolution 연산을 하면 26x26이 된다.&lt;/p&gt;&lt;p&gt;그러나 실제로 이미지 분석을 하는 경우에는, 2차원이지만 채널이 여러개라 마치 3차원 같은 입력데이터(ex-RGB)가 들어오게 된다. 이런 경우는 2차원 Convolution을 만들어 적용하는데, &lt;strong&gt;&lt;div&gt;입력값의 채널 개수만큼 커널 채널을 만들어주는 것&lt;/div&gt;&lt;/strong&gt;이 포인트이다. 즉, 입력값의 채널 개수와 커널의 채널 개수가 같아야한다.&lt;/p&gt;&lt;p&gt;입력 채널 개수와 커널 채널 개수를 같게하여 합성곱연산을 수행하면, 각 채널의 연산들이 모두 더해지기 때문에 채널이 1이 되어 출력 값의 크기는 (&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O_H,O_W,1)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 된다.&lt;/p&gt;&lt;p&gt;만약 출력 채널의 개수를 여러개로 만들고 싶다면 어떻게 하면 될까? &lt;strong&gt;&lt;div&gt;커널 텐서 자체의 개수를 여러개로 하면된다.&lt;/div&gt;&lt;/strong&gt;&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:34.765625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;2d-conv&quot; title=&quot;2d-conv&quot; src=&quot;/static/ff525d3ccb5398ffe9c8fed8422d5bec/2bef9/2d-conv.png&quot; srcSet=&quot;/static/ff525d3ccb5398ffe9c8fed8422d5bec/6f3f2/2d-conv.png 256w,/static/ff525d3ccb5398ffe9c8fed8422d5bec/01e7c/2d-conv.png 512w,/static/ff525d3ccb5398ffe9c8fed8422d5bec/2bef9/2d-conv.png 1024w,/static/ff525d3ccb5398ffe9c8fed8422d5bec/71c1d/2d-conv.png 1536w,/static/ff525d3ccb5398ffe9c8fed8422d5bec/a878e/2d-conv.png 2048w,/static/ff525d3ccb5398ffe9c8fed8422d5bec/de9fe/2d-conv.png 2146w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위와 같이 3차원 커널의 개수를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O_C&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개 사용하면, 출력도 3차원 텐서로 얻을 수 있다. 이것이 오늘날의 CNN 연산에서 사용하는 Convolution 연산의 기본적인 형태이다.&lt;/p&gt;&lt;h2 id=&quot;convolution-연산의-역전파-이해하기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#convolution-%EC%97%B0%EC%82%B0%EC%9D%98-%EC%97%AD%EC%A0%84%ED%8C%8C-%EC%9D%B4%ED%95%B4%ED%95%98%EA%B8%B0&quot; aria-label=&quot;convolution 연산의 역전파 이해하기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Convolution 연산의 역전파 이해하기&lt;/h2&gt;&lt;p&gt;Convolution 연산도 선형변환인 것은 마찬가지이다. 따라서 기존의 역전파 방식과 동일한 방식으로 진행한다.&lt;/p&gt;&lt;p&gt;다만, Convolution 연산은 &lt;strong&gt;커널이 모든 입력 데이터에 공통으로 적용&lt;/strong&gt;되므로, &lt;strong&gt;&lt;div&gt;역전파를 계산할 때(Convolution 연산에 미분할 때)에도 Convolution 연산이 나오게 된다.&lt;/div&gt;&lt;/strong&gt;&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;msub&gt;&lt;mo&gt;∫&lt;/mo&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;R&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/msup&gt;&lt;/msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mo&gt;∫&lt;/mo&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;R&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/msup&gt;&lt;/msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo mathvariant=&quot;normal&quot; lspace=&quot;0em&quot; rspace=&quot;0em&quot;&gt;′&lt;/mo&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
\frac{\partial}{\partial x}[f*g](x) &amp;amp;= \frac{\partial}{\partial x}\int_{\mathbb{R}^d}f(y)g(x-y)dy \\
&amp;amp;= \int_{\mathbb{R}^d}f(y)\frac{\partial g}{\partial x}(x-y)dy \\
&amp;amp;= [f*g&amp;#x27;](x)
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:6.666779999999999em;vertical-align:-3.0833899999999996em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.5833899999999996em&quot;&gt;&lt;span style=&quot;top:-5.58339em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∗&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-0.9480500000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.0833899999999996em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.5833899999999996em&quot;&gt;&lt;span style=&quot;top:-5.58339em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;margin-right:0.44445em;position:relative;top:-0.0011249999999999316em&quot;&gt;∫&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:-0.36453em&quot;&gt;&lt;span style=&quot;top:-1.7880500000000001em;margin-left:-0.44445em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathbb mtight&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7820285714285713em&quot;&gt;&lt;span style=&quot;top:-2.786em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9119499999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;margin-right:0.44445em;position:relative;top:-0.0011249999999999316em&quot;&gt;∫&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:-0.36453em&quot;&gt;&lt;span style=&quot;top:-1.7880500000000001em;margin-left:-0.44445em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathbb mtight&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7820285714285713em&quot;&gt;&lt;span style=&quot;top:-2.786em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9119499999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3714399999999998em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-0.9480500000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.3714399999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∗&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.801892em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;′&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.0833899999999996em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\partial x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 시그널 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;g(x-y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 적용되어, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;mo mathvariant=&quot;normal&quot; lspace=&quot;0em&quot; rspace=&quot;0em&quot;&gt;′&lt;/mo&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;g&amp;#x27;&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.946332em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.751892em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;′&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대한 합성곱연산이 된다.&lt;/p&gt;&lt;p&gt;이를 그림으로 이해해보자.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:36.32812499999999%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA8klEQVQY03WQv08CMRTH7z/0n3JkMS5OLOjkQowDYSEMakwcHECIMAnhCOf17tq+9tq+tj5SY4DId3gvTT/v1zeLB3LOfSzXxtgQojHI2qr1umJSaeR1hQDxWFlKgfAYJajewyvjSvLN7PkKjJEgyqppDTpr0NpEImLij4o5l73+02pd5F+T8e1Fa4MQYrPdrYpSSGgawbkAUEpp7/1pMSh99/gmlPO6/Hy5dhg1gZrQ4MNeKZ6u/XuzD/Pp/P1+0B/NmMxT0wM4nr85UcDGne7lzTBnC3rRyHheGQAUxXdd12TD3/6UHZJBZP8+/jub/n4A7EqR+MwM/UIAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;conv-back&quot; title=&quot;conv-back&quot; src=&quot;/static/0386fb10537155b0c2254b1d47fa3a36/2bef9/conv-back.png&quot; srcSet=&quot;/static/0386fb10537155b0c2254b1d47fa3a36/6f3f2/conv-back.png 256w,/static/0386fb10537155b0c2254b1d47fa3a36/01e7c/conv-back.png 512w,/static/0386fb10537155b0c2254b1d47fa3a36/2bef9/conv-back.png 1024w,/static/0386fb10537155b0c2254b1d47fa3a36/71c1d/conv-back.png 1536w,/static/0386fb10537155b0c2254b1d47fa3a36/a878e/conv-back.png 2048w,/static/0386fb10537155b0c2254b1d47fa3a36/96191/conv-back.png 2176w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;순전파에서, 입력값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_3&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 출력값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;으로 갈 때에는 커널의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_3&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O_2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 갈 때에는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O_3&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 갈 때에는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;파트를 사용했다.반대로, &lt;strong&gt;&lt;div&gt;역전파 과정에서 들어온 미분값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;δ&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\delta_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03785em&quot;&gt;δ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03785em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 순전파 시 적용된 가중치, 즉 커널의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_3&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 곱해져서 Gradient로 전달된다.&lt;/div&gt;&lt;/strong&gt;&lt;/p&gt;&lt;p&gt;이와 같이 역전파 단계에서는, 입력 단계에서 곱해졌던 커널(파트)을 통해 입력값에 Gradient를 전달한다. 그러면 입력값에서 각각의 커널들에는 어떻게 전달될까?&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:34.375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA7klEQVQY04VQPU8CQRC93+dfsqO10cR/QCgt0NLCEGhosCBcoQkkgBAg3LJ3t183s7PM3qkhAfFlszvZvHlv3iThBAgwmsy0cVx7Cv8iaR6iyDVGt7tDqaCQ83HvzoeglJJ5AeiZ4GvUBTVFcqpUlmWnO/icrtezYa9zw825lPOvzSo7OOc0a5sIa53WBgB/neNtrWHn5a5wajHpP0Rnray1RFfH/s7sKU0/3p9en/vjndhgJlRuDtILL2QlA/3RzBl+cou31uPt/ctimzKbVwhIQIiEF5w5wz7LhBBMq5fMCvEg+qqCBnQ2N/8g4hGUy5G88+pfZwAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;conv-back2&quot; title=&quot;conv-back2&quot; src=&quot;/static/f190f66baf3d9a2e86d2f62560178522/2bef9/conv-back2.png&quot; srcSet=&quot;/static/f190f66baf3d9a2e86d2f62560178522/6f3f2/conv-back2.png 256w,/static/f190f66baf3d9a2e86d2f62560178522/01e7c/conv-back2.png 512w,/static/f190f66baf3d9a2e86d2f62560178522/2bef9/conv-back2.png 1024w,/static/f190f66baf3d9a2e86d2f62560178522/71c1d/conv-back2.png 1536w,/static/f190f66baf3d9a2e86d2f62560178522/a878e/conv-back2.png 2048w,/static/f190f66baf3d9a2e86d2f62560178522/d6602/conv-back2.png 2222w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02778em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_3&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 전달된 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;δ&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\delta_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03785em&quot;&gt;δ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03785em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;값은, 기존에 적용되었던 커널 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_3&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;으로 다시 전달되므로 커널 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_3&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;δ&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\delta_1x_3&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03785em&quot;&gt;δ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03785em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;으로 업데이트된다. 마찬가지로 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;도 각각 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;δ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;δ&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\delta_2x_3,\delta_3x_3&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03785em&quot;&gt;δ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03785em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03785em&quot;&gt;δ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03785em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;으로 업데이트된다.&lt;/p&gt;&lt;p&gt;그런데, 각각의 커널들은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X_3&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;뿐만 아니라 다른 입력값에 대해서도 동일하게 적용되었으므로, 각 커널로 들어오는 Gradient를 모두 합해줘야한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:35.9375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAABDElEQVQY02VRS0oEMRScC3oCD+AN3MxRBEFE0Z2484MbEUHQjeCAow2jNN1tf6aTfi/fl7TJtI0zWoRapOq9KpJJvwYi97bItNL9P5RlBYCMcSGEMYaIwuVk0LyPLLXZO75s0kx31fvjfvAEnxDSOZ/lhZQyz4u6aRCF1vp3eABHcXByVb0my3R+e7TFeQcAn2m2KKoOsGUBHFAAIuNdqPk3+fDsvvlIjSifr6fGklIq5MSWzllyFA9Za39q+xWGFUF+eJolFzezpGRQcQ7OWm9MeIzI1vajc8Ba7UHovs53ptu7py+Lu3YJVikvpZOSEAmFd25jGBHrumlb5lbCuNqHnhTbRfKbgePX0DcKC4851u0CcQAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;conv-back3&quot; title=&quot;conv-back3&quot; src=&quot;/static/2e1ff91f73f9dc557fbda9c15439c913/2bef9/conv-back3.png&quot; srcSet=&quot;/static/2e1ff91f73f9dc557fbda9c15439c913/6f3f2/conv-back3.png 256w,/static/2e1ff91f73f9dc557fbda9c15439c913/01e7c/conv-back3.png 512w,/static/2e1ff91f73f9dc557fbda9c15439c913/2bef9/conv-back3.png 1024w,/static/2e1ff91f73f9dc557fbda9c15439c913/71c1d/conv-back3.png 1536w,/static/2e1ff91f73f9dc557fbda9c15439c913/a878e/conv-back3.png 2048w,/static/2e1ff91f73f9dc557fbda9c15439c913/895f3/conv-back3.png 2158w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;δ&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;δ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;δ&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{\partial\mathcal{L}}{\partial w_1} = \delta_1x_1 + \delta_2x_2 + \delta_3x_3&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.20744em;vertical-align:-0.8360000000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.3139999999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8360000000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03785em&quot;&gt;δ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03785em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03785em&quot;&gt;δ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03785em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03785em&quot;&gt;δ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03785em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이렇게 각 커널로 들어오는 모든 gradient들을 다 더하면 결국 convolution 연산과 동일한 연산이 이루어진다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/munder&gt;&lt;msub&gt;&lt;mi&gt;δ&lt;/mi&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;j&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{\partial\mathcal{L}}{\partial w_i} = \sum_j\delta_jx_{i+j-1}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.20744em;vertical-align:-0.8360000000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.3139999999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.02691em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8360000000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.463782em;vertical-align:-1.413777em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.8723309999999997em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.413777em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03785em&quot;&gt;δ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03785em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.311664em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05724em&quot;&gt;j&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;</content:encoded></item><item><title><![CDATA[딥러닝 베이직 - 05.CNN을 활용한 컴퓨터비전 기초]]></title><description><![CDATA[Computer Vision Applicaiton by 최성준 교수님, BoostCamp AI Tech 3주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/17_computer_vision_with_cnn/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/17_computer_vision_with_cnn/</guid><pubDate>Wed, 03 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;computer-vision-applications&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#computer-vision-applications&quot; aria-label=&quot;computer vision applications permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Computer Vision Applications&lt;/h1&gt;&lt;h2 id=&quot;semantic-segmentation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#semantic-segmentation&quot; aria-label=&quot;semantic segmentation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Semantic Segmentation&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Semantic Segmentation&lt;/code&gt;&lt;/strong&gt;은 이미지의 모든 픽셀을 어떤 레이블에 속하는지에 따라 (추정)분류하는 문제이다. &lt;code&gt;Dense Classification&lt;/code&gt;, &lt;code&gt;Per Pixel Classification&lt;/code&gt; 등으로 불리기도 한다.&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;fully-convolutional-network&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#fully-convolutional-network&quot; aria-label=&quot;fully convolutional network permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Fully Convolutional Network&lt;/h3&gt;&lt;p&gt;기존의 CNN에서 Dense Layer를 없애고, Conv Layer로 바꾼다. &lt;strong&gt;1x1 커널을 만들어 입력층의 크기와 동일한 output을 출력&lt;/strong&gt;해준다. 이를 &lt;strong&gt;&lt;code&gt;Convolutionalization&lt;/code&gt;&lt;/strong&gt;이라고 부른다.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;그런데 Convolutionalization을 하더라도 파라미터 숫자와 상태는 변함이 없다. 그냥 마지막 층만 Dense Layer에서 Conv Layer로 바뀌었을 뿐이다. Semantic Segmentation에서는 어떤 장점이 있길래 이걸 하는걸까?&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이를 이해하기 위해서는 &lt;strong&gt;Semantic Segmentation과 기존의 Classification 문제의 차이점&lt;/strong&gt;부터 짚어야한다. 기존의 분류문제는 해당 이미지가 &lt;strong&gt;&amp;#x27;어떤것&amp;#x27;&lt;/strong&gt;을 의미하는지만 알면 되었지만, &lt;div&gt;Semantic Segmentation은 이미지에서 &lt;strong&gt;&amp;#x27;어떤 것&amp;#x27;&lt;/strong&gt;이 &lt;strong&gt;&amp;#x27;어떤 위치&amp;#x27;&lt;/strong&gt;에 있는지까지 알아내야한다. &lt;/div&gt;&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;따라서, flatten으로 위치정보를 모두 없애버리고, &lt;strong&gt;&amp;#x27;어떤 것&amp;#x27;인지 분류만 하는&lt;/strong&gt; &lt;code&gt;전연결계층(Fully Connected Layer)&lt;/code&gt;은 적합하지 않다. 반면 같은 파라미터를 그대로 남겨놓음에도 불구하고 &lt;strong&gt;위치정보를 가지고 있는&lt;/strong&gt; &lt;strong&gt;&lt;code&gt;FCN(Fully Convolutional Network)&lt;/code&gt;&lt;/strong&gt;은 출력값으로 &lt;strong&gt;heat map을 제공&lt;/strong&gt;한다. 즉, &lt;strong&gt;&amp;quot;해당 이미지에서 A라는 물체가 &amp;#x27;어디에&amp;#x27; 있는가?&amp;quot;에 대한 정보를 제공&lt;/strong&gt;해준다.&lt;/p&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;다만, [Conv+Pooling] 연산을 여러번 거치게 되면서 feature-map이 줄어들게 되는데, 이를 원래의 dense pixels로 바꿔줄 필요가 있다. 이를 &lt;strong&gt;&lt;code&gt;Deconvolution&lt;/code&gt;&lt;/strong&gt; 또는 &lt;strong&gt;&lt;code&gt;Upsampling&lt;/code&gt;&lt;/strong&gt;이라고 부른다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;일종의 Conv 역연산을 수행한다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;줄어들었던 Space를 다시 원래의 Space로 복원&lt;/strong&gt;한다. 그러나 엄밀히 말하면 완전한 복원은 아니다. Conv 연산에서 더해진 값들이 정확히 각 픽셀마다 얼마였는지 알 수 없기 때문이다.&lt;/li&gt;&lt;li&gt;Conv 연산의 결과인 &lt;strong&gt;feature-map의 바깥쪽, 사이사이에 패딩을 적절히 주어서 역연산하여 원하는 크기로 늘린다.&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;자세한 설명은 다음 글을 참고한다.
&lt;a href=&quot;https://m.blog.naver.com/PostView.nhn?blogId=laonple&amp;amp;logNo=220958109081&amp;amp;proxyReferer=https:%2F%2Fwww.google.com%2F&quot;&gt;[Part Ⅶ. Semantic Segmentation] 3. FCN [1] - 라온피플 머신러닝 아카데미 -&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;detection&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#detection&quot; aria-label=&quot;detection permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Detection&lt;/h2&gt;&lt;p&gt;앞에서 언급했던 Semantic Segmentation처럼 이미지에서 물체의 &amp;#x27;위치&amp;#x27;까지 찾아내는 것이다. 그러나 픽셀단위로 분류하는 것이 아니라, Bounding Box 단위로 어떤 물체인지 인식한다. 즉, 이미지에서 물체를 박스 단위로 포착하고, 해당 물체가 어떤 것인지 클래스 분류까지 진행한다.&lt;/p&gt;&lt;h1 id=&quot;-7&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-7&quot; aria-label=&quot; 7 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;r-cnn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#r-cnn&quot; aria-label=&quot;r cnn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;R-CNN&lt;/h3&gt;&lt;p&gt;Detection 분야의 시초라고 할 수 있는 모델링이 &lt;strong&gt;&lt;code&gt;R-CNN&lt;/code&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;먼저 이미지에서 독립적으로 약 2,000개의 랜덤한 &lt;code&gt;Region(Regional Proposal)&lt;/code&gt;을 뽑는다. 이 region들은 크기도 각각 다르다. 각 region들을 똑같은 크기로 맞춰주고 CNN을 통해 고정된 길이의 feature vector를 추출한다.  이후, 각 region마다 linear SVM으로 분류를 진행하여 어떤 물체가 있는지 확인한다.&lt;/p&gt;&lt;h1 id=&quot;-8&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-8&quot; aria-label=&quot; 8 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;자세한 설명은 다음 글을 참고하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://jaehyeongan.github.io/2019/10/10/R-CNN/&quot;&gt;R-CNN(Regions with CNN features) 논문 리뷰&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-9&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-9&quot; aria-label=&quot; 9 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;그러나, R-CNN의 문제는, 약 2천번의 독립추출이 필요하고, 그렇게 뽑아낸 &lt;strong&gt;각 Region마다 feature vector 추출과 Classification이 들어간다&lt;/strong&gt;는 것이다. 자연스레 연산량이 굉장히 많아지게 된다. 이 때문에 한 이미지를 처리하는 데에 약 1분이 걸리게 된다.&lt;/p&gt;&lt;h1 id=&quot;-10&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-10&quot; aria-label=&quot; 10 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;sppnet&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#sppnet&quot; aria-label=&quot;sppnet permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;SPPNet&lt;/h3&gt;&lt;p&gt;R-CNN의 높은 연산량 문제를 해결하기 위해 &lt;strong&gt;&lt;code&gt;SPPNet&lt;/code&gt;&lt;/strong&gt;이 만들어졌다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;R-CNN&lt;/code&gt;&lt;ul&gt;&lt;li&gt;image → &lt;strong&gt;crop/warp&lt;/strong&gt; → conv layers → fc layers → output&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;SPPNet&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;image → conv layers → &lt;strong&gt;spatial pyramid pooling&lt;/strong&gt; → fc layers → output&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이미지 전체에 대해 &lt;strong&gt;&lt;div&gt;한번의 CNN만 수행&lt;/div&gt;&lt;/strong&gt;하여 Convolution Feature Map을 만든 이후, Bounding Box 내부에 위치하는 텐서만 뜯어서 사용한다.&lt;/p&gt;&lt;h1 id=&quot;-11&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-11&quot; aria-label=&quot; 11 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;fast-r-cnn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#fast-r-cnn&quot; aria-label=&quot;fast r cnn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Fast R-CNN&lt;/h3&gt;&lt;p&gt;&lt;code&gt;SPPNet&lt;/code&gt;은 CNN 횟수를 한번으로 줄였지만, 알맞은 Bounding Box를 찾아 Classification하는 데에 사용되는 수많은 연산을 해결하지는 못했다. 이를 개선한 버전이 &lt;strong&gt;&lt;code&gt;Fast R-CNN&lt;/code&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;h1 id=&quot;-12&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-12&quot; aria-label=&quot; 12 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;Fast R-CNN의 기본적인 원리는 SPPNet과 비슷하다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;이미지를 입력받고, 약 2천개의 Bounding Box를 생성한다.&lt;/li&gt;&lt;li&gt;한번의 CNN으로 Convolutional feature map을 생성한다.&lt;/li&gt;&lt;li&gt;각 region마다 ROI(Region Of Interest) pooling으로 fixed length feature를 추출한다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;div&gt;두개의 출력값으로 class분류와 bounding-box regressor를 동시에 업데이트&lt;/div&gt;&lt;/strong&gt;한다.&lt;/li&gt;&lt;/ol&gt;&lt;h1 id=&quot;-13&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-13&quot; aria-label=&quot; 13 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;faster-r-cnn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#faster-r-cnn&quot; aria-label=&quot;faster r cnn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Faster R-CNN&lt;/h3&gt;&lt;p&gt;지금까지는 Bounding Box를 뽑아내는 방식이 &lt;code&gt;selective search&lt;/code&gt;, 임의의 박스를 뽑는 알고리즘에 불과했지만, 이제는 이 박스를 뽑는 방법(Region Proposal Network, RPN)까지 학습시키기로 한다. &lt;code&gt;Region Proposal Network&lt;/code&gt;와 &lt;code&gt;Fast R-CNN&lt;/code&gt;을 합친 것이 &lt;strong&gt;&lt;code&gt;Faster R-CNN&lt;/code&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;h1 id=&quot;-14&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-14&quot; aria-label=&quot; 14 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:55.859375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;RPN&quot; title=&quot;RPN&quot; src=&quot;/static/3d112d2bb274b5e032a616bea5423fae/2bef9/RPN.png&quot; srcSet=&quot;/static/3d112d2bb274b5e032a616bea5423fae/6f3f2/RPN.png 256w,/static/3d112d2bb274b5e032a616bea5423fae/01e7c/RPN.png 512w,/static/3d112d2bb274b5e032a616bea5423fae/2bef9/RPN.png 1024w,/static/3d112d2bb274b5e032a616bea5423fae/71c1d/RPN.png 1536w,/static/3d112d2bb274b5e032a616bea5423fae/599ea/RPN.png 1926w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-15&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-15&quot; aria-label=&quot; 15 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;RPN은 Fully Conv Network를 활용하는데, Conv 과정에서 커널이 각 영역을 찍으며 미리 정해진 박스 크기를 의미하는 &lt;strong&gt;&lt;code&gt;Anchor Box(Predefiend region size)&lt;/code&gt;&lt;/strong&gt;를 정해둔다. &lt;strong&gt;Anchor Box에 해당 영역에 물체가 포함되어있는지 Classification 작업을 수행하고, 동시에 Bounding Box의 크기를 조절하는 Bouding Box Regression도 진행&lt;/strong&gt;한다.&lt;/p&gt;&lt;h1 id=&quot;-16&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-16&quot; aria-label=&quot; 16 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;yolov1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#yolov1&quot; aria-label=&quot;yolov1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;YOLO(v1)&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;YOLO(You Only Look Once)&lt;/code&gt;&lt;/strong&gt;는 기존의 Faster R-CNN보다 훨씬 빠르다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;Bounding box와 class probailities를 동시에 예측&lt;/strong&gt;하며, box를 샘플링하는 과정을 거치지 않는다.&lt;/p&gt;&lt;h1 id=&quot;-17&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-17&quot; aria-label=&quot; 17 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이미지를 cell들로 이루어진 grid로 나누고, 각각의 cell은 B개의 바운딩박스를 예측한다. 5개의 바운딩 박스 중 물체가 담겨있지 않은 쓸모없는 박스를 버리고, 즉시 C개의 class probailities를 분류한다. 따라서 &lt;strong&gt;바운딩 박스를 찾는 것과 class를 분류하는것이 동시에 작동하므로 훨씬 더 빠르게 작동&lt;/strong&gt;된다.&lt;/p&gt;&lt;h1 id=&quot;-18&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-18&quot; aria-label=&quot; 18 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;지금까지 언급된 CNN의 발전과정들을 풀이해놓은 글을 참고해보자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://junklee.tistory.com/35&quot;&gt;Image Detection 방법론: RCNN, SPPnet, FastRCNN, FasterRCNN&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[딥러닝 베이직 - 04.CNN 기초와 역사 훑어보기]]></title><description><![CDATA[CNN & Modern CNN by 최성준 교수님, BoostCamp AI Tech 3주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/16_cnn_basic/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/16_cnn_basic/</guid><pubDate>Wed, 03 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;convolution-neural-network-basic&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#convolution-neural-network-basic&quot; aria-label=&quot;convolution neural network basic permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Convolution Neural Network Basic&lt;/h1&gt;&lt;h2 id=&quot;원리&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9B%90%EB%A6%AC&quot; aria-label=&quot;원리 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;원리&lt;/h2&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;CNN&lt;/code&gt;&lt;/strong&gt;은 &lt;code&gt;합성곱 계층(convolution layer)&lt;/code&gt;과 &lt;code&gt;풀링 계층(pooling layer)&lt;/code&gt;, 그리고 &lt;code&gt;전연결 계층(fully-connected layer)&lt;/code&gt;으로 이루어져 있는 신경망으로, 이미지 등을 처리하는데에 자주 사용되는 모델이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;합성곱 계층, 풀링 계층 : feature를 추출한다.&lt;/li&gt;&lt;li&gt;전연결 계층 : 분류/회귀 문제에 대해 decision making한다(출력값을 만든다)&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;그러나 모델의 파라미터가 너무 많아지면 Generalization 이슈가 생기기 때문에, &lt;strong&gt;최근에는 전연결계층을 없애려고 하는 추세&lt;/strong&gt;이다.&lt;/p&gt;&lt;div&gt;&lt;p&gt;아래에서 풀이하는 CNN은 편의상 편향(bias)을 설명하지 않았지만, 실제로는 편향도 고려하여 연산에 추가하여야 한다는 것을 유의하자.&lt;/p&gt;&lt;/div&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;stride&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#stride&quot; aria-label=&quot;stride permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Stride&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Stride&lt;/code&gt;&lt;/strong&gt;는 &amp;#x27;넓게 걷는다&amp;#x27;라는 뜻으로, &lt;strong&gt;매번 커널을 찍을때 얼마나 이동할것인가&lt;/strong&gt;를 의미한다. stride가 클수록 output의 사이즈는 작아지게 된다.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;padding&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#padding&quot; aria-label=&quot;padding permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Padding&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;패딩(Padding)&lt;/code&gt;&lt;/strong&gt;은 입력 값의 가장자리를 커널로 찍기 위해 &lt;strong&gt;덧대는 공간&lt;/strong&gt;이다. 일반적으로 해당 패딩 위치에 0을 넣는 제로패딩을 사용한다.&lt;/p&gt;&lt;p&gt;커널의 크기를 어떻게 하느냐에 따라 패딩의 크기가 달라지기도 한다. 예를 들어, 3x3 커널에서는 패딩을 1칸만 추가해도 가장자리를 찍을 수 있겠지만, 5x5 커널에서는 모서리를 커널의 중심으로 두고 가장자리에 커널을 찍기위해 패딩을 2칸은 추가해야할 것이다.&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;convolution-파라미터-개수-구하기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#convolution-%ED%8C%8C%EB%9D%BC%EB%AF%B8%ED%84%B0-%EA%B0%9C%EC%88%98-%EA%B5%AC%ED%95%98%EA%B8%B0&quot; aria-label=&quot;convolution 파라미터 개수 구하기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Convolution 파라미터 개수 구하기&lt;/h2&gt;&lt;h3 id=&quot;두-계층-간의-파라미터-개수&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%91%90-%EA%B3%84%EC%B8%B5-%EA%B0%84%EC%9D%98-%ED%8C%8C%EB%9D%BC%EB%AF%B8%ED%84%B0-%EA%B0%9C%EC%88%98&quot; aria-label=&quot;두 계층 간의 파라미터 개수 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;두 계층 간의 파라미터 개수&lt;/h3&gt;&lt;p&gt;만약 (H,W,C)가 (40,50,128)인 입력값을 convolution연산하여 (40,50,64)의 출력값을 만들었다면, 그 사이에 들어가있는 파라미터는 몇개나 될까?&lt;/p&gt;&lt;ul&gt;&lt;li&gt;패딩은 1, 스트라이드는 1, 커널의 크기는 3x3이라고 한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;커널의 채널 크기는 입력값의 채널수와 같으므로, 커널은 (3,3,128)이 될 것이다. 커널의 갯수는 출력값의 채널 수와 같아야하므로 64개가 될 것이다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;패딩, 스트라이드 등은 파라미터 개수와 무관하다&lt;/strong&gt;. 파라미터 개수는 커널 (내부) 값들의 개수이므로, 3x3x128x64 = 73,728개가 된다.&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;fully-connected-layer가-있다면&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#fully-connected-layer%EA%B0%80-%EC%9E%88%EB%8B%A4%EB%A9%B4&quot; aria-label=&quot;fully connected layer가 있다면 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Fully Connected Layer가 있다면&lt;/h3&gt;&lt;p&gt;위에서 Fully Connected Layer가 있다면 파라미터 개수가 너무 많아져 Generalization 이슈가 생길 수 있다고 했는데, 일반적인 Convolution에 비해 왜 파라미터 개수가 많아질까?&lt;/p&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:46.09375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAJCAIAAAC9o5sfAAAACXBIWXMAABYlAAAWJQFJUiTwAAABTElEQVQoz0WR2XLCMAxF+f9/61Pp0A4UAolpnU3eHS8xvUnK4MzY8pWOpMi7x7pmVpeWz/V1vje45tMXjAL9c1+kKFqX48HZtD/mj2OSGp5HKY8djui9I3KX0zT0ujobIRT/1ZfvYC1dL57VTip3q5LSZ5YO56kb4wu2zv3wTmijretJSqV6Eu1A1jrtvLJWG6uc9yF4p72Tb+9jTwu/wCEEIkFESioppRCShByGwRgzDKPAXalxJO89ra6qUR2lF9x2PeqsWVBZYweBYl3fG2NhKK29n5AHXmN0jNM/nHNu2B3prbVwIkgp9OtjTKUUeOd5hkFCGGsjuGmCsMBlXagZ4rIhFCp4NAIdPJSc8vYoULZRbfZuO55qcW5pnnN+u9Wct1V1ZYxhIlNY0s0b91y7umFrHE9rTXSVUkaKbYSYE74QI8qvf/Ei0d0fNpIEUSx0hggAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;alex-net&quot; title=&quot;alex-net&quot; src=&quot;/static/cc2c71eb351b5bedae137db467b26ac4/2bef9/alex-net.png&quot; srcSet=&quot;/static/cc2c71eb351b5bedae137db467b26ac4/6f3f2/alex-net.png 256w,/static/cc2c71eb351b5bedae137db467b26ac4/01e7c/alex-net.png 512w,/static/cc2c71eb351b5bedae137db467b26ac4/2bef9/alex-net.png 1024w,/static/cc2c71eb351b5bedae137db467b26ac4/71c1d/alex-net.png 1536w,/static/cc2c71eb351b5bedae137db467b26ac4/ec09f/alex-net.png 1916w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;위의 그림은 AlexNet의 파라미터 개수를 나타낸다. 잘 보면 Fully Connected Layer인 Dense Layer에 진입하는 순간 파라미터의 개수가 엄청나게 커지는 것을 볼 수 있다. 이는 &lt;strong&gt;전연결시에 모든 입력 노드(값)들을 다음 계층의 노드들과 연결하여 파라미터를 생성하기 때문&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;이와 다르게 Covolution 연산에서 사용하는 파라미터, 커널은 모든 채널과 input에 대해 동일하게 동작하는 &lt;code&gt;공유 파라미터(shared parameter)&lt;/code&gt;이다. 따라서 전연결 계층에 비해 파라미터 개수가 훨씬 줄어들게 되어 결과적으로 Generalization이 잘 되는 특징을 가진다. 이 때문에 &lt;strong&gt;최근 CNN의 트렌드는 앞단의 Convolution Layer를 깊게 쌓고, 뒤쪽의 Fully Connected Layer를 최대한 줄이는 것&lt;/strong&gt;이다. 이를 통해 네트워크의 깊이는 깊어지지만, 파라미터 개수는 오히려 줄어들어, 적은 연산으로도 더 뛰어난 성능을 낼 수 있게 된다.&lt;/p&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;1x1-convolution&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#1x1-convolution&quot; aria-label=&quot;1x1 convolution permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;1x1 Convolution&lt;/h2&gt;&lt;p&gt;가끔 Convolution 연산을 보다보면, 1x1의 커널로 연산하는 경우가 있다. 1x1로 연산한다면 H와 W의 값을 조정할수도 없을텐데 왜 하는것일까?&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;차원축소(Demension Reduction)&lt;/code&gt;&lt;/strong&gt; 때문이다. 기존의 H,W라는 공간차원은 그대로 유지한 채, &lt;strong&gt;채널만 줄이기 위해 사용&lt;/strong&gt;한다. 이를 통해 &lt;strong&gt;&lt;div&gt;깊이(depth)를 늘리는 도중에 파라미터의 개수를 감소&lt;/div&gt;&lt;/strong&gt;시킬수 있게 된다.&lt;/p&gt;&lt;p&gt;아주 자주 사용하는 테크닉으로, 예시로 bottleneck architecture 등이 있다.&lt;/p&gt;&lt;h1 id=&quot;-7&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-7&quot; aria-label=&quot; 7 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;cnn의-변천사&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#cnn%EC%9D%98-%EB%B3%80%EC%B2%9C%EC%82%AC&quot; aria-label=&quot;cnn의 변천사 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;CNN의 변천사&lt;/h2&gt;&lt;h3 id=&quot;alexnet&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#alexnet&quot; aria-label=&quot;alexnet permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;AlexNet&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://kr.nvidia.com/content/tesla/pdf/machine-learning/imagenet-classification-with-deep-convolutional-nn.pdf&quot;&gt;Imagenet Classification with Deep Convolutional NN 논문&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;http://www.image-net.org/challenges/LSVRC/#:~:text=The%20ImageNet%20Large%20Scale%20Visual,image%20classification%20at%20large%20scale.&amp;amp;text=Another%20motivation%20is%20to%20measure,indexing%20for%20retrieval%20and%20annotation.&quot;&gt;ILSVRC&lt;/a&gt;에서 2012년 우승한 모델이다.&lt;/p&gt;&lt;p&gt;여러 필터 중 하나로 11x11을 사용했는데, 이렇게 하면 하나의 커널이 볼 수 있는 영역은 커지지만, 상대적으로 더 많은 파라미터가 필요하므로 그렇게 좋은 선택은 아니다.&lt;/p&gt;&lt;p&gt;모델의 핵심 포인트는 다음과 같다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Rectified Linear Unit(ReLU) 활성함수 사용&lt;ul&gt;&lt;li&gt;선형 모델의 프로퍼티를 보존 :  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;&amp;gt;&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&amp;gt;0&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.5782em;vertical-align:-0.0391em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;&amp;gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이면 gradient가 선형모델과 동일하다.&lt;/li&gt;&lt;li&gt;경사하강법으로 최적화하기 쉽다.&lt;/li&gt;&lt;li&gt;일반화 성능이 좋다(실험결과)&lt;/li&gt;&lt;li&gt;&lt;strong&gt;기울기 소실 문제(Vanishing Gradient Problem)을 해결하였다.&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;GPU Inplementation (2개의 GPU 사용) - 당시 하드웨어 기술의 메모리 부족으로 인하여 2개의 GPU를 사용하였다.&lt;/li&gt;&lt;li&gt;Local response normalization, Overlapping pooling&lt;ul&gt;&lt;li&gt;지금은 잘 사용되지 않는다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;Data augmentation&lt;/li&gt;&lt;li&gt;Dropout(p=0.5)&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-8&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-8&quot; aria-label=&quot; 8 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;vggnet&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#vggnet&quot; aria-label=&quot;vggnet permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;VGGNet&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://arxiv.org/pdf/1409.1556.pdf%20http://arxiv.org/abs/1409.1556.pdf&quot;&gt;Very Deep Convolutional Networks For Large-Scale Image Recognition 논문&lt;/a&gt;&lt;/p&gt;&lt;p&gt;2014년도 ILSVRC에서 준우승한 모델이다.&lt;/p&gt;&lt;p&gt;모델의 핵심포인트는 다음과 같다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;3x3 필터(스트라이드 1)을 사용하고 depth를 늘렸다.&lt;/li&gt;&lt;li&gt;전연결계층에서 1x1 convolution 연산을 사용했다.&lt;/li&gt;&lt;li&gt;Dropout(p=0.5)&lt;/li&gt;&lt;li&gt;계층 수에 따라 VGG16, VGG19로 나뉜다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;모든 레이어에서 &lt;strong&gt;필터를 3x3으로 통일&lt;/strong&gt;하였다. 이것은 어떤 의미를 가질까?&lt;/p&gt;&lt;p&gt;필터가 크다는 것은, 순전파 과정에서 하나의 convolution feature 값을 얻기 위해 고려할 수 있는 입력의 spatial dimension, 즉 &lt;strong&gt;&lt;code&gt;수용영역(Receptive Field)&lt;/code&gt;&lt;/strong&gt;이 커진다는 말이다.&lt;/p&gt;&lt;p&gt;만약 3x3 필터를 두 번 사용하게 된다면, 결국 5x5를 한번 사용하는것과 &lt;strong&gt;&lt;div&gt;수용영역 차원에서는 동일&lt;/div&gt;&lt;/strong&gt;하다. 3x3을 하나의 셀로 축소시킨것을 다시 3x3개 모아서 축소시키기 때문이다.&lt;/p&gt;&lt;p&gt;그러나 파라미터 개수에는 큰 차이가 있다. 입력, 출력의 채널이 각각 (128,128)이라고 생각해보자.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;128&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;128&lt;/mn&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;128&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;128&lt;/mn&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mrow&gt;&lt;mn mathvariant=&quot;bold&quot;&gt;294&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mn mathvariant=&quot;bold&quot;&gt;912&lt;/mn&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mn&gt;5&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;5&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;128&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;128&lt;/mn&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mrow&gt;&lt;mn mathvariant=&quot;bold&quot;&gt;409&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mn mathvariant=&quot;bold&quot;&gt;600&lt;/mn&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
3\times 3\times 128\times 128 +3\times3\times 128\times 128 &amp;amp;=\bold{294,912}\\
5\times 5\times 128\times 128 &amp;amp;= \bold{409,600}
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.0000000000000004em;vertical-align:-1.2500000000000002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7500000000000002em&quot;&gt;&lt;span style=&quot;top:-3.91em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;3&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;3&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;8&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;8&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;3&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;3&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;8&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;8&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.41em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;8&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;8&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2500000000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.7500000000000002em&quot;&gt;&lt;span style=&quot;top:-3.91em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord mathbf&quot;&gt;9&lt;/span&gt;&lt;span class=&quot;mord mathbf&quot;&gt;4&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathbf&quot;&gt;9&lt;/span&gt;&lt;span class=&quot;mord mathbf&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord mathbf&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.41em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;4&lt;/span&gt;&lt;span class=&quot;mord mathbf&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord mathbf&quot;&gt;9&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathbf&quot;&gt;6&lt;/span&gt;&lt;span class=&quot;mord mathbf&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord mathbf&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2500000000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;따라서 &lt;strong&gt;&lt;div&gt;같은 수용영역을 커버하면서 파라미터의 개수를 줄이기 위해서&lt;/div&gt;&lt;/strong&gt;는 3x3으로 계층을 한번 더 쌓는것이 낫다.&lt;/p&gt;&lt;p&gt;이러한 이유로 최근에는 대부분 필터의 크기가 3x3, 커봤자 7x7을 넘지 않는다.&lt;/p&gt;&lt;h1 id=&quot;-9&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-9&quot; aria-label=&quot; 9 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;googlenet&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#googlenet&quot; aria-label=&quot;googlenet permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;GoogLeNet&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://arxiv.org/pdf/1409.4842.pdf&quot;&gt;Christian et al, &amp;quot;Going Deeper with Convolutions&amp;quot;, CVPR, 2015 논문&lt;/a&gt;&lt;/p&gt;&lt;p&gt;2014년도 ILSVRC에서 우승한 모델이다.&lt;/p&gt;&lt;p&gt;22단으로 이루어져 있으며, 비슷한 네트워크가 전체 네트워크 내부에 여러번 들어가 있다. 이를 &lt;code&gt;Network in Network,NiN&lt;/code&gt; 구조라고 부른다.&lt;/p&gt;&lt;p&gt;입력값이 Convolution 연산으로 들어가기전에 1x1 Conv를 거치는데, 이를 &lt;strong&gt;&lt;code&gt;Inception Block&lt;/code&gt;&lt;/strong&gt;이라고 하며 GoogLeNet 모델에서 가장 중요한 부분이다. Inception Block을 이용하여 파라미터 숫자를 줄일 수 있게 된다. 왜 그럴까?&lt;/p&gt;&lt;p&gt;1x1 Convolution이 &lt;strong&gt;채널방향으로(channel-wise) 차원을 축소하는 효과가 있기 때문&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:44.921875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAJCAIAAAC9o5sfAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA+klEQVQoz31Ri27DIAzM/39k10nbtNCBCe9XYBe8qlErzYoCyHdnn72MMW4/8nJ5+/j80noj0kqpVQgiQqq1FmMspdRa930fHL3zuTAi51wqMAVPoEMIQPMdl5SztTbGlFLGvz+RWVtrjYLAjT7AZzLKZWvMTXyvotVaWjPvV0fUn8jGWOecsRbPEOKD7JxVEu7SUTtLIQD7qwwz3gMcuVUAvPfsEHfw0S90YS3NGPdY+IDStm2HwHQW7wjYs0c3DrPEH464QD+T/wmYQll8aEcRxXNlaCPNeufoc6Z9muKJzNE8Fob0gmbWVUipkDgzsbvC25v7e20Hcr/ddg5D0I3ETQAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;inception-block&quot; title=&quot;inception-block&quot; src=&quot;/static/3df49d2429692e67e279477ed6d3f1a6/2bef9/inception-block.png&quot; srcSet=&quot;/static/3df49d2429692e67e279477ed6d3f1a6/6f3f2/inception-block.png 256w,/static/3df49d2429692e67e279477ed6d3f1a6/01e7c/inception-block.png 512w,/static/3df49d2429692e67e279477ed6d3f1a6/2bef9/inception-block.png 1024w,/static/3df49d2429692e67e279477ed6d3f1a6/71c1d/inception-block.png 1536w,/static/3df49d2429692e67e279477ed6d3f1a6/a4f52/inception-block.png 2024w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;입력과 출력의 채널도 같고 , 수용영역도 같지만 1x1 convolution을 중간에 추가하는 것만으로  &lt;strong&gt;파라미터의 개수는 1/3보다 더 줄어들었다.&lt;/strong&gt;&lt;/p&gt;&lt;h1 id=&quot;-10&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-10&quot; aria-label=&quot; 10 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;resnet&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#resnet&quot; aria-label=&quot;resnet permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;ResNet&lt;/h3&gt;&lt;p&gt;&lt;a href=&quot;https://openaccess.thecvf.com/content_cvpr_2016/papers/He_Deep_Residual_Learning_CVPR_2016_paper.pdf&quot;&gt;Deep Residual Learning, CVPR, 2016 논문&lt;/a&gt;&lt;/p&gt;&lt;p&gt;2015년도 ILSVRC에서 우승한 모델이다.&lt;/p&gt;&lt;p&gt;신경망의 깊이가 깊을수록, 학습하기는 더 어려워진다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;오버피팅은 일반적으로 파라미터 개수가 너무 많을 경우 일어난다.&lt;/li&gt;&lt;li&gt;그러나 깊이가 깊은 경우에는 오버피팅이 일어나지 않음에도 불구하고 학습에러에 비해 테스트에러가 훨씬 크게 나온다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;그래서 ResNet은 &lt;strong&gt;&lt;code&gt;identity map(skip connection)&lt;/code&gt;&lt;/strong&gt;이라는 것을 추가했다. &lt;strong&gt;입력값으로 들어온 x를 convolution layer 한 층의 출력값에 더해주는 것&lt;/strong&gt;이다. 이렇게 만든 block을 &lt;strong&gt;&lt;code&gt;Residual Block&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;/p&gt;&lt;p&gt;기존의 신경망은 입력값 x를 목표 출력값 y로 매핑하는 함수 H(x)를 얻는것이 목표였는데, ResNet은 F(x)+x를 최소화시키는것이 목표다. 이 때 x는 정해진 상수이므로 F(x)를 최소화해야한다. F(x) = H(x)-x, 즉 &lt;strong&gt;&lt;code&gt;잔차(Residual)&lt;/code&gt;&lt;/strong&gt;이므로 잔차를 최소화한다고 하여 ResNet이라는 이름이 붙게 된다.&lt;/p&gt;&lt;p&gt;자세한 설명은 아래 링크를 참고하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://itrepo.tistory.com/36&quot;&gt;(7) ResNet (Residual Connection)&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://blueskyvision.tistory.com/644&quot;&gt;[CNN 알고리즘들] ResNet의 구조&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-11&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-11&quot; aria-label=&quot; 11 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이렇게 학습시키게 되면, &lt;strong&gt;이미 배웠던 내용 x를 제외한 차이점(f(x))만을 학습&lt;/strong&gt;하면 되므로 학습이 더 잘이루어지게 된다.&lt;/p&gt;&lt;h1 id=&quot;-12&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-12&quot; aria-label=&quot; 12 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:53.515625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;resnet&quot; title=&quot;resnet&quot; src=&quot;/static/2b040dacd75a63af148e4583fc01cb0d/2bef9/resnet.png&quot; srcSet=&quot;/static/2b040dacd75a63af148e4583fc01cb0d/6f3f2/resnet.png 256w,/static/2b040dacd75a63af148e4583fc01cb0d/01e7c/resnet.png 512w,/static/2b040dacd75a63af148e4583fc01cb0d/2bef9/resnet.png 1024w,/static/2b040dacd75a63af148e4583fc01cb0d/71c1d/resnet.png 1536w,/static/2b040dacd75a63af148e4583fc01cb0d/a2ef2/resnet.png 1970w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-13&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-13&quot; aria-label=&quot; 13 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이 때, 위의 Simple Shortcut과 같은 방식으로 진행하려면 두번의 conv를 거치고 난 뒤 BN까지 진행하고 나서의 (채널)차원이 기존의 입력값 x와 같아야한다. 만약 차원이 다르다면, 이를 맞춰주기 위해 1x1 Conv로 채널을 맞춰주게 된다. 이를 &lt;code&gt;Projected Shortcut&lt;/code&gt;이라고 한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Projected Shortcut이 자주 사용되지는 않으며, 일반적으로 &lt;code&gt;Simple Shortcut&lt;/code&gt;을 많이 사용한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;또, ResNet은 Batch Normalization이 Convolution 뒤에 일어난다는 특징이 있다. 다만, Batch Normalization을 ReLu 뒤에 넣는것이 더 학습효과가 좋다는 논쟁도 있다. 논문에서는 위의 이미지와 같은 순서로 수행한다.&lt;/p&gt;&lt;h1 id=&quot;-14&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-14&quot; aria-label=&quot; 14 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:42.96875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAJCAIAAAC9o5sfAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA90lEQVQoz3WRTU+EMBCG+f+/SW8mJh6MazjsQdhgUWmLLtDS7w4O2xVZV95DMzOdZz7arOuH9w9Uwxlv25ZS1jRUjuO0CKYtZWsHALz310kBooe4CSNmnaOM4gxKKbRTPKJZfdJdwfKDGZW7LH2GY4xCiENVUcawuUXazrwXij++7G8f+HMZjDPWhhD+gYdBkLr+Oh7xeuaNmWGp2a7Ib+5ZXk4hWmvXe/2O7YOvCOGcG62NsRDnhwLlTEHLu6duT7Qcl3XOMPwIHSmlOTVM7qJBSR/DOp6M7Dr1r2Dzw7K6fnslpOv7VCUVcs7p0/B4uMtRU5EU/wZg+Qz/nfjEkQAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;bottleneck&quot; title=&quot;bottleneck&quot; src=&quot;/static/b61cfcca81088c506ebb9c749c67ebb0/2bef9/bottleneck.png&quot; srcSet=&quot;/static/b61cfcca81088c506ebb9c749c67ebb0/6f3f2/bottleneck.png 256w,/static/b61cfcca81088c506ebb9c749c67ebb0/01e7c/bottleneck.png 512w,/static/b61cfcca81088c506ebb9c749c67ebb0/2bef9/bottleneck.png 1024w,/static/b61cfcca81088c506ebb9c749c67ebb0/71c1d/bottleneck.png 1536w,/static/b61cfcca81088c506ebb9c749c67ebb0/a878e/bottleneck.png 2048w,/static/b61cfcca81088c506ebb9c749c67ebb0/d8fc6/bottleneck.png 2052w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-15&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-15&quot; aria-label=&quot; 15 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;3x3 Conv를 수행할 때, Conv의 파라미터 개수는 [3x3x입력채널x출력채널]이다. 그렇다면 &lt;strong&gt;1x1 Conv로 입력채널을 줄여서 넣고, 출력 후에는 채널을 늘릴 수도 있을 것&lt;/strong&gt;이다. 이를 &lt;strong&gt;&lt;code&gt;Bottleneck architecture&lt;/code&gt;&lt;/strong&gt;라고 한다.&lt;/p&gt;&lt;h1 id=&quot;-16&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-16&quot; aria-label=&quot; 16 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;densenet&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#densenet&quot; aria-label=&quot;densenet permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;DenseNet&lt;/h3&gt;&lt;p&gt;ResNet의 Residual Block에서 x값을 더해(addition)주면 두 값이 섞이게 된다. 그래서 이 과정을 &lt;strong&gt;&lt;code&gt;연쇄적으로 잇는 것(concatenation)&lt;/code&gt;으로 대체&lt;/strong&gt;한다.&lt;/p&gt;&lt;h1 id=&quot;-17&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-17&quot; aria-label=&quot; 17 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:42.578125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;densenet&quot; title=&quot;densenet&quot; src=&quot;/static/b4ffde9ec486da5eab705dcffab895ac/2bef9/densenet.png&quot; srcSet=&quot;/static/b4ffde9ec486da5eab705dcffab895ac/6f3f2/densenet.png 256w,/static/b4ffde9ec486da5eab705dcffab895ac/01e7c/densenet.png 512w,/static/b4ffde9ec486da5eab705dcffab895ac/2bef9/densenet.png 1024w,/static/b4ffde9ec486da5eab705dcffab895ac/71c1d/densenet.png 1536w,/static/b4ffde9ec486da5eab705dcffab895ac/a878e/densenet.png 2048w,/static/b4ffde9ec486da5eab705dcffab895ac/57f7d/densenet.png 2088w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-18&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-18&quot; aria-label=&quot; 18 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;문제는 concatenation 하면 할수록 &lt;strong&gt;채널수가 점점 기하급수적으로 커진다&lt;/strong&gt;는데에 있다. 채널 수가 많아지면 Conv 연산 시 파라미터수가 자연스레 많아지므로, 이를 해결하기 위해 중간에 채널숫자를 줄여주는 &lt;strong&gt;&lt;code&gt;Transition Block&lt;/code&gt;&lt;/strong&gt;을 끼워넣는다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Dense Block&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;각 층은 모든 이전층의 feature map을 concatenate한다.&lt;/li&gt;&lt;li&gt;채널의 수는 기하급수적으로 커지게 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Transition Block&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;BatchNorm → 1x1 Conv → 2x2 AvgPooling&lt;/li&gt;&lt;li&gt;채널 차원을 축소한다.(Dimension Reduction)&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;</content:encoded></item><item><title><![CDATA[딥러닝 베이직 - 03.최적화 기법]]></title><description><![CDATA[Optimization by 최성준 교수님, BoostCamp AI Tech 3주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/14_optimization/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/14_optimization/</guid><pubDate>Tue, 02 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;optimization&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#optimization&quot; aria-label=&quot;optimization permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Optimization&lt;/h1&gt;&lt;h2 id=&quot;용어-정리&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9A%A9%EC%96%B4-%EC%A0%95%EB%A6%AC&quot; aria-label=&quot;용어 정리 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;용어 정리&lt;/h2&gt;&lt;h3 id=&quot;generalization&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#generalization&quot; aria-label=&quot;generalization permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Generalization&lt;/h3&gt;&lt;p&gt;많은 경우에 일반화 성능을 높이는 것이 학습의 목적이 된다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Generalization Gap&lt;/code&gt;&lt;/strong&gt;은 &lt;strong&gt;학습오류(Training Error)와 테스트오류(Test Error)사이의 간극&lt;/strong&gt;을 의미한다. 좋은 &lt;strong&gt;&lt;code&gt;일반화 성능(Generalization Performace)&lt;/code&gt;&lt;/strong&gt;이 보장되어있다는 것은 이 갭이 작다는 말이다. 이는 곧, &lt;strong&gt;해당 모델의 (테스트 데이터에 대한) 성능이 학습데이터와 비슷할 것&lt;/strong&gt;임을 나타낸다.&lt;/p&gt;&lt;p&gt;그런데, 학습 데이터의 자체가 좋지 않아 모델의 성능이 떨어지는 경우가 있을수도 있다. 이 경우에도 학습 데이터와 테스트 데이터의 차이가 적다면 일반화 성능은 좋다고 표현한다. &lt;strong&gt;&lt;div&gt;따라서, 일반화 성능이 좋다고해서 모델의 성능이 좋다는 것을 의미하지는 않는다.&lt;/div&gt;&lt;/strong&gt;&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;under-fitting-vs-over-fitting&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#under-fitting-vs-over-fitting&quot; aria-label=&quot;under fitting vs over fitting permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Under-fitting vs over-fitting&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;오버피팅(Overfitting, 과적합)&lt;/code&gt;&lt;/strong&gt;은 학습 데이터에 대해서는 잘 동작하지만, 테스트 데이터에는 잘 동작하지 않는 것을 의미한다.&lt;/p&gt;&lt;p&gt;반대로, &lt;strong&gt;&lt;code&gt;언더피팅(Underfitting, 과소적합)&lt;/code&gt;&lt;/strong&gt;은 모델의 구성이 너무 간단하거나 학습을 너무 적게 해서 학습 데이터도 잘 맞추지 못하는 것을 의미한다.&lt;/p&gt;&lt;p&gt;이는 개념적인 이야기라서, 해당 모델이 오버피팅/언더피팅임을 판단하는 기준은 절대적이지 않다.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;cross-validation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#cross-validation&quot; aria-label=&quot;cross validation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Cross-validation&lt;/h3&gt;&lt;p&gt;데이터를 분리하여, 학습데이터는 학습만, 테스트 데이터는 모델의 성능을 테스트하는 validation만 수행한다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;교차검증(Cross-validation)&lt;/code&gt;&lt;/strong&gt;은 학습데이터와 테스트 데이터를 분리하여 검증하는 기술을 의미하는데, 가장 대표적인 것으로 &lt;code&gt;k-fold validation&lt;/code&gt;이 있다.&lt;/p&gt;&lt;p&gt;&lt;code&gt;K-fold validation&lt;/code&gt;은 테스트 데이터를 제외한 학습 데이터를 k개로 나누어서, k-1개로 학습을 시키고 나머지 한 개로 validation을 하는 기법이다. 이 때 성능을 제대로 검증하기 위해 validation으로 사용하는 하나를 매번 바꿔가며 테스트해본다.&lt;/p&gt;&lt;p&gt;신경망은 최적의 &lt;code&gt;파라미터&lt;/code&gt;를 구하기 위한 학습과정에서 &lt;strong&gt;&lt;code&gt;하이퍼파라미터(Hyperparameter)&lt;/code&gt;&lt;/strong&gt;가 필요하다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;파라미터&lt;/code&gt;&lt;ul&gt;&lt;li&gt;가중치와 bias, CNN의 필터 등&lt;/li&gt;&lt;li&gt;구하고자 하는 값들&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;하이퍼파라미터&lt;/code&gt;&lt;ul&gt;&lt;li&gt;학습률(learning rate), 손실함수(loss function), 신경망의 크기 등&lt;/li&gt;&lt;li&gt;직접 정하는 값들&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;처음에는 어떤 하이퍼파라미터가 좋은지 모르기 때문에, 교차검증으로 하이퍼파라미터를 먼저 찾고나서 모든 학습 데이터를 사용해 파라미터를 구한다.&lt;/p&gt;&lt;div&gt;&lt;p&gt;테스트 데이터는 절대로 학습에 사용하지 않는다. 테스트 데이터는 오로지 테스트만을 위해 사용할 뿐, 학습에 사용되는 것은 그 자체로 cheating으로 본다.&lt;/p&gt;&lt;/div&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;bias-variance-tradeoff&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#bias-variance-tradeoff&quot; aria-label=&quot;bias variance tradeoff permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Bias-variance tradeoff&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:834px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:103.90625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;bias-and-variance&quot; title=&quot;bias-and-variance&quot; src=&quot;/static/c2e4baf47a183d89b3c049ff71eeae68/5d72a/bias-and-variance.png&quot; srcSet=&quot;/static/c2e4baf47a183d89b3c049ff71eeae68/6f3f2/bias-and-variance.png 256w,/static/c2e4baf47a183d89b3c049ff71eeae68/01e7c/bias-and-variance.png 512w,/static/c2e4baf47a183d89b3c049ff71eeae68/5d72a/bias-and-variance.png 834w&quot; sizes=&quot;(max-width: 834px) 100vw, 834px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;분산(Variance)&lt;/code&gt;&lt;/strong&gt;은 &lt;strong&gt;input을 넣었을 때 얼마나 output이 일관적으로 나오는가&lt;/strong&gt;를 나타낸다. 일반적으로 Variance가 낮은 모델들은 간단한 모델들이다. 이에 비해 Variance가 높은 모델들은 input에 따라 output이 비교적 크게 바뀌므로, 오버피팅될 가능성이 높다.&lt;/p&gt;&lt;p&gt;사격에서의 탄착군 형성을 생각해보자. &lt;strong&gt;&lt;div&gt;중요한 것은 과녁 중심을 얼마나 맞히느냐가 아니라 탄착지점이 얼마나 모여있느냐&lt;/div&gt;&lt;/strong&gt;이다. 탄착군이 모여있으면, 가늠쇠를 조정하여 탄착군을 과녁 중앙으로 shift하면 된다.&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;편향(Bias)&lt;/code&gt;&lt;/strong&gt;은 분산과 관계없이, &lt;strong&gt;평균값이 목표치(True Target), 즉 정답에 얼마나 접근했는가&lt;/strong&gt;를 나타낸다. 탄착군이 형성되지 않더라도, 점들의 평균치가 과녁 중앙에 가까이 갔다면 편향이 낮은 것이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mtext&gt;Given &lt;/mtext&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;{&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;msubsup&gt;&lt;mo stretchy=&quot;false&quot;&gt;}&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/msubsup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mspace width=&quot;1em&quot;&gt;&lt;/mspace&gt;&lt;mtext&gt;Where&lt;/mtext&gt;&lt;mspace width=&quot;1em&quot;&gt;&lt;/mspace&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;ϵ&lt;/mi&gt;&lt;mspace width=&quot;1em&quot;&gt;&lt;/mspace&gt;&lt;mtext&gt;and&lt;/mtext&gt;&lt;mspace width=&quot;1em&quot;&gt;&lt;/mspace&gt;&lt;mi&gt;ϵ&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\textrm{Given}\ D = \{(x_i,t_i)\}^N_{i=1}, \quad\textrm{Where}\quad t = f(x) + \epsilon \quad \textrm{and} \quad  \epsilon \sim \mathcal{N}(0,\sigma^2)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;Given&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1413309999999999em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913309999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:1em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;Where&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:1em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;ϵ&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:1em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord text&quot;&gt;&lt;span class=&quot;mord textrm&quot;&gt;and&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:1em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;ϵ&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1141079999999999em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.14736em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;Bias and Variance의 트레이드오프 관계는 오래된 주제이다. 아래의 식은 학습데이터에 노이즈가 껴 있음을 가정하고 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo fence=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo fence=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo fence=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo fence=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo fence=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo fence=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo fence=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo fence=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;ϵ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
\mathbb{E}\Big[(t-\hat{f})^2\Big] &amp;amp;=  \mathbb{E}\Big[(t-f+f-\hat{f})^2\Big]\\
&amp;amp;= \dots\\
&amp;amp;= \mathbb{E}\Big[(f-\mathbb{E}[\hat{f}])^2\Big] + \mathbb{E}\Big[(\mathbb{E}[\hat{f}]-\hat{f})^2\Big]+\mathbb{E}[\epsilon]
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:5.7000399999999996em;vertical-align:-2.6000199999999998em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.1000199999999998em&quot;&gt;&lt;span style=&quot;top:-5.10002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size2&quot;&gt;[&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9578799999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.26344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.08332999999999999em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size2&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.31em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.5000000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.6000199999999998em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:3.1000199999999998em&quot;&gt;&lt;span style=&quot;top:-5.10002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size2&quot;&gt;[&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9578799999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.26344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.08332999999999999em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size2&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.31em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.5000000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size2&quot;&gt;[&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9578799999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.26344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.08332999999999999em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size2&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size2&quot;&gt;[&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9578799999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.26344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.08332999999999999em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9578799999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.26344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.08332999999999999em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size2&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;ϵ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.6000199999999998em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;t&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.61508em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 True Target, 즉  목표치(정답)을 의미한다.&lt;/li&gt;&lt;li&gt;첫 식은 cost를 의미한다.&lt;/li&gt;&lt;li&gt;마지막 식에서 각 항은 왼쪽부터 차례대로 bias&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8141079999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, variance, noise를 의미한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Cost를 최소화하기 위해 총 세 파트, [bias&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8141079999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, variance, noise]를 줄여야하는데, 이 중 &lt;strong&gt;bias와 variance는 한 쪽을 줄이면 다른 한 쪽이 늘어나는 관계&lt;/strong&gt;이다. 근본적으로 &lt;strong&gt;둘 모두를 줄이기는 힘들다&lt;/strong&gt;는 fundamental이 존재한다.&lt;/p&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;bootstrapping&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#bootstrapping&quot; aria-label=&quot;bootstrapping permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Bootstrapping&lt;/h3&gt;&lt;p&gt;&amp;#x27;부츠 끈을 들어서 하늘을 날겠다&amp;#x27;라는 의미의 bootstrap라는 어원에서 유래하였다. &amp;#x27;지금 가지고 있는 것으로 어떻게든 활용하여 해결하겠다&amp;#x27;라고 볼 수 있는데, 통계학에서는 가설검증이나 metric 계산 이전에 랜덤샘플링을 적용하는 방법을 일컫는다.&lt;/p&gt;&lt;p&gt;서로 다른 학습데이터를 사용했을 때 모델의 예측은 차이가 나게 되는데, 이 &lt;strong&gt;예측의 통계(또는 일치도)를 보고 전체적인 모델의 불확실성(uncertainty)를 예측&lt;/strong&gt;하고자 할 때 &lt;strong&gt;&lt;code&gt;부트스트래핑&lt;/code&gt;&lt;/strong&gt;을 사용한다.&lt;/p&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;여러가지로 활용될 수 있지만, 일반적으로는 &lt;strong&gt;학습데이터가 고정되어있을 때 서브샘플링을 통해 여러 학습데이터를 만들고 여러 모델과 metric을 만들어 사용하겠다&lt;/strong&gt;는 말이다.&lt;/p&gt;&lt;h1 id=&quot;-7&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-7&quot; aria-label=&quot; 7 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;bagging-and-boosting&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#bagging-and-boosting&quot; aria-label=&quot;bagging and boosting permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Bagging and Boosting&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:54.296875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;bootstrapping-boosting&quot; title=&quot;bootstrapping-boosting&quot; src=&quot;/static/2416def990d58b51efd7154386d6f566/2bef9/bootstrapping-boosting.png&quot; srcSet=&quot;/static/2416def990d58b51efd7154386d6f566/6f3f2/bootstrapping-boosting.png 256w,/static/2416def990d58b51efd7154386d6f566/01e7c/bootstrapping-boosting.png 512w,/static/2416def990d58b51efd7154386d6f566/2bef9/bootstrapping-boosting.png 1024w,/static/2416def990d58b51efd7154386d6f566/78958/bootstrapping-boosting.png 1320w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Bagging&lt;/code&gt;&lt;/strong&gt;(&lt;strong&gt;B&lt;/strong&gt;ootstrapping &lt;strong&gt;agg&lt;/strong&gt;regat&lt;strong&gt;ing&lt;/strong&gt;)은 학습데이터가 고정되어있을 때, 학습데이터를 다 사용하는 것이 아니라 랜덤샘플링한 여러개로 만들어 여러개의 모델을 만든 뒤 이 모델들의 output을 평균낸다는 것을 의미한다. 부트스트래핑과 관련있는 말이며, &lt;strong&gt;&lt;code&gt;앙상블(Ensemble)&lt;/code&gt;&lt;/strong&gt;이라고 불리기도 한다.&lt;/p&gt;&lt;p&gt;직관적으로는 모든 데이터를 학습데이터로 썼을 때 가장 좋은 성능의 모델이 나올 것 같지만, 실제로는 &lt;strong&gt;오히려 데이터들을 서브샘플링 한 뒤 여러 모델의 예측을 평균낸 것이 더 성능이 좋은 경우가 많다.&lt;/strong&gt; &lt;/p&gt;&lt;h1 id=&quot;-8&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-8&quot; aria-label=&quot; 8 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Boosting&lt;/code&gt;&lt;/strong&gt;은  조금 다른 방법이다.&lt;/p&gt;&lt;p&gt;학습데이터를 여러 개로 샘플링한 뒤, 하나의 모델을 가지고 각 샘플링 데이터들을 차례대로 학습시킨다. 해당 모델이 제대로 예측하지 못하는 데이터들에 대해서는, &lt;strong&gt;해당 데이터들만 따로 학습시키는 모델을 새로 만든다&lt;/strong&gt;. 이 개개의 모델들을 &lt;code&gt;weak learner&lt;/code&gt;라고 부른다.&lt;/p&gt;&lt;p&gt;이렇게 여러개의 모델을 만든 뒤 sequential하게 합쳐서 하나의 strong learner를 만드는 것을 부스팅이라고 한다. &lt;/p&gt;&lt;h1 id=&quot;-9&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-9&quot; aria-label=&quot; 9 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;practical-gradient-descent-method&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#practical-gradient-descent-method&quot; aria-label=&quot;practical gradient descent method permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Practical Gradient Descent Method&lt;/h2&gt;&lt;h3 id=&quot;경사하강법---배치-방식에-따라&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B2%BD%EC%82%AC%ED%95%98%EA%B0%95%EB%B2%95---%EB%B0%B0%EC%B9%98-%EB%B0%A9%EC%8B%9D%EC%97%90-%EB%94%B0%EB%9D%BC&quot; aria-label=&quot;경사하강법   배치 방식에 따라 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;경사하강법 - 배치 방식에 따라&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;Stochastic Gradient Descent(SGD)&lt;/code&gt;&lt;ul&gt;&lt;li&gt;한번에 하나의 샘플로만 gradient를 구해서 업데이트하는 방식&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;Mini-batch Gradient Descent&lt;/code&gt;&lt;ul&gt;&lt;li&gt;한번에 데이터의 일부(일반적으로 배치 사이즈)로 gradient를 구해서 업데이트하는 방식&lt;/li&gt;&lt;li&gt;가장 많이 사용하는 방식&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;Batch Gradient Descent&lt;/code&gt;&lt;ul&gt;&lt;li&gt;한번에 모든 데이터로 gradient를 구해서 업데이트하는 방식&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-10&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-10&quot; aria-label=&quot; 10 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;배치사이즈는-얼마나-커야하는가&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%B0%B0%EC%B9%98%EC%82%AC%EC%9D%B4%EC%A6%88%EB%8A%94-%EC%96%BC%EB%A7%88%EB%82%98-%EC%BB%A4%EC%95%BC%ED%95%98%EB%8A%94%EA%B0%80&quot; aria-label=&quot;배치사이즈는 얼마나 커야하는가 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;배치사이즈는 얼마나 커야하는가?&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:41.40625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAICAIAAAB2/0i6AAAACXBIWXMAABYlAAAWJQFJUiTwAAAAyklEQVQY022R2w6EIAxE/f/fdB8UVqMJAqUt7gBeNus2PJB0Du0M3X5WrrU/alkWY6wxJoaQYyyHqLW6bx2nlFWfPBGBRFfmWaZJ13WvY244cSlh/iFVNRHhPFs3HGIU6M7hzYBUEnMwnIkOa6e7run0NKu5qA8+Z07cgiA8/ZwsmnGqIF/JBe/HYbTGUjyygee3tdM8O+du+E++lcf+WFVE2j/g4gPe9HQtDxg/0fevYRihrtgBI6NEadu82zauYcJ/KcH6HCPa6QOuWdfrf0sHHQAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;batch-size&quot; title=&quot;batch-size&quot; src=&quot;/static/b94e18aecbf25326bbcf37977bf2e862/2bef9/batch-size.png&quot; srcSet=&quot;/static/b94e18aecbf25326bbcf37977bf2e862/6f3f2/batch-size.png 256w,/static/b94e18aecbf25326bbcf37977bf2e862/01e7c/batch-size.png 512w,/static/b94e18aecbf25326bbcf37977bf2e862/2bef9/batch-size.png 1024w,/static/b94e18aecbf25326bbcf37977bf2e862/71c1d/batch-size.png 1536w,/static/b94e18aecbf25326bbcf37977bf2e862/b8471/batch-size.png 2016w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;몇몇 논문의 실험 결과 &lt;strong&gt;배치사이즈가 너무 커지면&lt;/strong&gt; &lt;code&gt;Sharp Minimum&lt;/code&gt;에 빠지게 되고, 이는 곧 테스트 데이터에 대한 좋지 않은 성능으로 연결된다는 것이 밝혀졌다.&lt;/p&gt;&lt;p&gt;반대로, 배치사이즈가 작을수록 noise의 영향력이 커지므로 Sharp Minimum에서 탈출할 확률이 높다.&lt;/p&gt;&lt;h1 id=&quot;-11&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-11&quot; aria-label=&quot; 11 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;해당 실험을 다룬 &lt;a href=&quot;https://arxiv.org/pdf/1609.04836.pdf&quot;&gt;On Large-batch Training for Deep Learning : Generalization Gap and Sharp Minima, 2017&lt;/a&gt; 논문을 참조해보자.&lt;/p&gt;&lt;p&gt;자세한 설명은 아래 글을 참조한다.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://iamseungjun.tistory.com/3&quot;&gt;Small Batch Size in Deep Learning&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-12&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-12&quot; aria-label=&quot; 12 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;gradient-descent-methods&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#gradient-descent-methods&quot; aria-label=&quot;gradient descent methods permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Gradient Descent Methods&lt;/h3&gt;&lt;ol&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;(Stochastic) Gradient Descent&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;가장 기본적인 경사하강법&lt;/li&gt;&lt;li&gt;learning rate를 정하는 것이 어렵다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Momentum&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;관성처럼, &lt;strong&gt;이전 배치에서의 방향성을 어느정도 유지&lt;/strong&gt;하는 방식&lt;/li&gt;&lt;li&gt;모멘텀 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;β&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\beta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05278em&quot;&gt;β&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;과 현재 gradient &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;g&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;g&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 합친 accumulation를 이용해 업데이트한다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;div&gt;gradient가 요동치더라도 이전의 관성이 남아있기 때문에 어느정도 학습이 잘 된다.&lt;/div&gt;&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Nesterov Accelerated Gradient(NAG)&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;Lookahead gradient를 계산한다.&lt;/li&gt;&lt;li&gt;기존의 모멘텀은 관성으로 인해 local minimum 부근에서 수렴(converging)하지 못하는 문제가 있었지만, NAG는 모멘텀으로 업데이트 시 발생하는 변화를 미리 체크한 뒤 모멘텀을 결정하므로 좀 더 minimum에 빠르게 수렴할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Adagrad&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;이전까지의 모멘텀방식과 다른 ADA(adaptive) 방식&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;G&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 지금까지의 gradient 제곱 합을 저장시키고, 역수에 집어넣는다. 따라서 많이 변한 파라미터는 적게 변화하게되고, 적게 변한 파라미터는 많이 변화하게 된다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;G&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 계속 커지기 때문에, 분모가 점점 커지므로 learning rate가 0에 수렴해 학습이 멈춘다는 문제가 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Adadelta&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;Adagrad에서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;G&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;G&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;G&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 계속 커지는 문제를 exponential moving average를 통해 해결하려는 방식&lt;ul&gt;&lt;li&gt;지수이동평균은 최근 데이터에 가중치를 부여하고, 과거의 데이터일수록 영향력이 작아지는 형태를 띈다.&lt;/li&gt;&lt;li&gt;단, Adadelta에서는 윈도우사이즈만큼의 local average를 구한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;시간을 기준으로 window를 잡아, 최근에 gradient가 많이 변했으면 적게 변화시키고, 적게 변화했으면 많이 변화시킨다.&lt;/li&gt;&lt;li&gt;윈도우 사이즈만큼의 gradient 정보를 들고 있어야하는데, GPT-3같은 천억개 수준의 파라미터를 가진 모델의 정보를 다 들고 있을 수 없다는 문제가 있다.&lt;/li&gt;&lt;li&gt;learning rate가 없다는 특징이 있는데, 이 때문에 변화를 주기가 어려우므로 잘 사용되지 않는다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;RMSprop&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;제프리 힌턴이 딥러닝 강의를 하다가 언급했던 경험을 바탕으로 구현한 방식&lt;/li&gt;&lt;li&gt;adadelta처럼 exponential moving average를 이용하지만, 대신 분자에 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;η&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\eta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;η&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라는 stepsize(learning rate)를 넣는다.&lt;/li&gt;&lt;li&gt;adagrad랑 adadelta랑 섞어놓은 듯한, 과거에 많이 사용되었던 방식이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;Adam&lt;/code&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;Adaptive와 Momentum 방식들을 합친 것으로, 가장 많이 사용하는 방식이다.&lt;/li&gt;&lt;li&gt;모멘텀 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;m&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;과 gradient 제곱 값들의 합 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;v&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 사용한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;ϵ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\epsilon&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;ϵ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;은 division by zero를 막기 위한 하이퍼파라미터인데, 아주 작은 값을 넣는다. 실제로는 이 값을 잘 조정하는것이 practical한 성능에 큰 영향을 주는 것으로 나타났다.&lt;ul&gt;&lt;li&gt;0이 아니면서 0에 얼마나 가깝게 만드는가?&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;더 자세한 설명은 다음 글을 참조하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;http://shuuki4.github.io/deep%20learning/2016/05/20/Gradient-Descent-Algorithm-Overview.html&quot;&gt;Gradient Descent Optimization Algorithms 정리&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-13&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-13&quot; aria-label=&quot; 13 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;regularization&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#regularization&quot; aria-label=&quot;regularization permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Regularization&lt;/h2&gt;&lt;p&gt;학습을 방해하도록 규제함으로써, 모델이 학습 데이터에만 피팅되는것이 아니라 테스트 데이터에서도 잘 동작하도록 만드는 것, 즉 오버피팅을 피하도록 하는 것을 &lt;strong&gt;&lt;code&gt;정규화(regularization)&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;/p&gt;&lt;h1 id=&quot;-14&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-14&quot; aria-label=&quot; 14 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;early-stopping&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#early-stopping&quot; aria-label=&quot;early stopping permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Early Stopping&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:42.96875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAJCAIAAAC9o5sfAAAACXBIWXMAABYlAAAWJQFJUiTwAAABCUlEQVQoz5VRh27FIAzM//9i1eTlLVYWw4wAoSZpmw6pUk/Isuw7jI+mfGBZZM65Ztt2nj/R7ORKenm9hbCW/+AUtxc6TKakNXIWOU+M5mX+MTyEYK11zmOCquaoYkaZ6G5TACje5r7NlBTnDnFKyTknpQSAdUeMsU7GZsqVoZTyPhJmRh1V2wcpUykW3hGC335Z0MS05b06T7MxBh/Fhbw/x64XbcfuRD6ZphzYYKkAwuQigzLZh2ptcyjxVlwDxQAWLEajjWFcjOOkdS3XDlildUzpnLztwJWQgtF7z7m49FeMj8eTEopAJbbwUZgg4RQTQoUYYvz2SegQfhu6EnZ3Putf10bb3wAaBAvC3ZXMLAAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;early-stopping&quot; title=&quot;early-stopping&quot; src=&quot;/static/3b12d5ca95b88ad5db61c41f7eaa0bcc/2bef9/early-stopping.png&quot; srcSet=&quot;/static/3b12d5ca95b88ad5db61c41f7eaa0bcc/6f3f2/early-stopping.png 256w,/static/3b12d5ca95b88ad5db61c41f7eaa0bcc/01e7c/early-stopping.png 512w,/static/3b12d5ca95b88ad5db61c41f7eaa0bcc/2bef9/early-stopping.png 1024w,/static/3b12d5ca95b88ad5db61c41f7eaa0bcc/71c1d/early-stopping.png 1536w,/static/3b12d5ca95b88ad5db61c41f7eaa0bcc/769f8/early-stopping.png 1924w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt; &lt;/p&gt;&lt;p&gt;validation 에러와 학습 에러의 차이가 날 정도로 학습하기 전에 학습을 멈추어 버리는것을 &lt;strong&gt;&lt;code&gt;Early stopping&lt;/code&gt;&lt;/strong&gt;이라고 한다.&lt;/p&gt;&lt;h1 id=&quot;-15&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-15&quot; aria-label=&quot; 15 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;parameter-norm-penalty&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#parameter-norm-penalty&quot; aria-label=&quot;parameter norm penalty permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Parameter Norm Penalty&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Parameter Norm Penalty&lt;/code&gt;&lt;/strong&gt;는 신경망의 파라미터가 너무 커지지 않도록 조절하는 것을 의미한다.&lt;/p&gt;&lt;p&gt;신경망 파라미터의 숫자들이 너무 커지지 않도록 하기 위해 파라미터 값들(의 제곱의 합)을 줄여준다.&lt;/p&gt;&lt;p&gt;파라미터의 크기가 크지않은, 즉 부드러운(smooth) 함수일수록 일반화성능(Generalization Performance)이 좋을것이라는 가정에 기초하고 있다.&lt;/p&gt;&lt;h1 id=&quot;-16&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-16&quot; aria-label=&quot; 16 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;data-augmentation&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#data-augmentation&quot; aria-label=&quot;data augmentation permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Data Augmentation&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:33.203125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA7klEQVQY022P3Y7EIAiF+/5vuTc7U2Nt/QEE0S7WzGYv9kQTFD7gbABQa72nxrzjBiREsmCMAY8suP/Tdp4XIi64E9TgIAWDVdW+yqO/AEtDotVuq0S9zylKSRm1dWb67HLLo9Xa6qsosXYDFgyIoqMzKJO9mypQZhHVvjtXCrTWHksjVybhjDGmZPyEC2DvOrhYWpRzBSBjm/U2O7ZXk3Z3SSUUiENK9Pt1XfysYzCc/jvEI+QYUo6FmGUZnoswJ7/791cIR6s0j/BvdlNtMV0JELCaParknLPey+Wcn+Pr9T6OkypbhfdHCOeCfwAi2ZhO8OwHkQAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;data-augmentation&quot; title=&quot;data-augmentation&quot; src=&quot;/static/6515fb5afafd6f54bc02b9749afeb1a6/2bef9/data-augmentation.png&quot; srcSet=&quot;/static/6515fb5afafd6f54bc02b9749afeb1a6/6f3f2/data-augmentation.png 256w,/static/6515fb5afafd6f54bc02b9749afeb1a6/01e7c/data-augmentation.png 512w,/static/6515fb5afafd6f54bc02b9749afeb1a6/2bef9/data-augmentation.png 1024w,/static/6515fb5afafd6f54bc02b9749afeb1a6/71c1d/data-augmentation.png 1536w,/static/6515fb5afafd6f54bc02b9749afeb1a6/8454b/data-augmentation.png 1798w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;일반적으로 데이터가 많으면 모델의 성능이 더 좋다. 거꾸로 말하면, 데이터가 없으면 딥러닝 모델은 다른 머신러닝 방법론에 비하여 열등하다.&lt;/p&gt;&lt;p&gt;그러나 데이터셋이 어느정도 커지게 되면, 딥러닝 모델은 기존의 머신러닝 방법론들이 따라갈 수 없는 표현력을 가지게 된다. 결국 관건은 데이터의 양이다.&lt;/p&gt;&lt;h1 id=&quot;-17&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-17&quot; aria-label=&quot; 17 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Data Augmentation&lt;/code&gt;&lt;/strong&gt;은 어떤 식으로든 내가 &lt;strong&gt;가지고 있는 데이터를 변형시켜서 데이터 풀을 늘리는 방법&lt;/strong&gt;을 의미한다. 단, 변화는 이미지의 레이블이 바뀌지 않는 한도 내에서 수행해야한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;예를 들어, MNIST 데이터에서 6을 변형시키겠다고 뒤집어 9로 만드는것은 잘못된 data augmentation이다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-18&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-18&quot; aria-label=&quot; 18 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;noise-robustness&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#noise-robustness&quot; aria-label=&quot;noise robustness permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Noise Robustness&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:48.4375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;noise-robustness&quot; title=&quot;noise-robustness&quot; src=&quot;/static/64d6a3b1a79fd167618faf95cbdedccd/2bef9/noise-robustness.png&quot; srcSet=&quot;/static/64d6a3b1a79fd167618faf95cbdedccd/6f3f2/noise-robustness.png 256w,/static/64d6a3b1a79fd167618faf95cbdedccd/01e7c/noise-robustness.png 512w,/static/64d6a3b1a79fd167618faf95cbdedccd/2bef9/noise-robustness.png 1024w,/static/64d6a3b1a79fd167618faf95cbdedccd/c929c/noise-robustness.png 1218w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Noise Robustness&lt;/code&gt;&lt;/strong&gt;는 입력데이터에 Noise를 집어넣는것이다. Data Augmentation과 다른점은, &lt;strong&gt;단순히 입력에만 Noise를 집어넣는 것이 아니라 weight에도 집어넣는 것&lt;/strong&gt;을 의미한다.&lt;/p&gt;&lt;p&gt;정확한 원리가 규명되지는 않았고, 실험결과 상 데이터 풀이 많아질 뿐 아니라, weight 값을 흔들어 모델의 성능에 도움이 된다고 밝혀졌다.&lt;/p&gt;&lt;h1 id=&quot;-19&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-19&quot; aria-label=&quot; 19 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;label-smoothing&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#label-smoothing&quot; aria-label=&quot;label smoothing permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Label Smoothing&lt;/h3&gt;&lt;p&gt;Data Augmentation과 비슷하지만, &lt;strong&gt;&lt;code&gt;Label Smoothing&lt;/code&gt;&lt;/strong&gt;은 랜덤한 학습 데이터를 두 개 뽑아서 조작하여 사용한다는 차이점이 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;Mix-up&lt;/code&gt;&lt;ul&gt;&lt;li&gt;두 데이터를 섞어서(blending) 사용한다.&lt;/li&gt;&lt;li&gt;ex - 강아지와 고양이 그림을 겹쳐서 섞음&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;CutMix&lt;/code&gt;&lt;ul&gt;&lt;li&gt;두 데이터를 섞는것이 아니라 파트별로 잘라붙인 하나의 데이터를 만든다.&lt;/li&gt;&lt;li&gt;ex - 강아지 사진의 머리부분과 고양이 사진의 몸부분을 잘라 붙인 사진을 만든다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이를 통해 &lt;strong&gt;decision boundary를 더 부드럽게 만들어준다&lt;/strong&gt;.&lt;/p&gt;&lt;h1 id=&quot;-20&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-20&quot; aria-label=&quot; 20 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;dropout&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#dropout&quot; aria-label=&quot;dropout permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Dropout&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Dropout&lt;/code&gt;&lt;/strong&gt;은 순전파의 각 과정에서 신경망의 가중치 중 특정 비율을 0으로 바꾸는 것을 의미한다.&lt;/p&gt;&lt;p&gt;이를 통해 각각의 뉴런들이 좀 더 robust한 feature를 담을 수 있다고 한다.&lt;/p&gt;&lt;h1 id=&quot;-21&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-21&quot; aria-label=&quot; 21 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;batch-normalization&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#batch-normalization&quot; aria-label=&quot;batch normalization permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Batch Normalization&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;Batch Normalization&lt;/code&gt;&lt;/strong&gt;은 &lt;strong&gt;각 레이어의 statistics(활성화값, 또는 출력값)를 정규화&lt;/strong&gt;시키는 것을 의미한다.&lt;/p&gt;&lt;p&gt;Internal Covariate Shift를 줄인다고 논문에서 이야기하는데, 최근의 많은 논문들은 이에 동의하지않아 논란이 있다.&lt;/p&gt;&lt;p&gt;그러나, 적용 시 일반적으로 성능이 많이 높아진다는 실험 결과가 있다. 특히, 층이 깊을수록 더 그렇다고 한다.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://arxiv.org/pdf/1803.08494.pdf&quot;&gt;Group Normalization, 2018 논문&lt;/a&gt;을 참고해보자.&lt;/p&gt;&lt;p&gt;풀이한 글은 아래를 참조하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://blog.lunit.io/2018/04/12/group-normalization/&quot;&gt;Group Normalization&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[딥러닝 베이직 - 01.딥러닝의 간략한 역사]]></title><description><![CDATA[딥러닝 기본용어 설명 및 Historical Review by 최성준 교수님, BoostCamp AI Tech 3주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/12_historical_review_of_dl/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/12_historical_review_of_dl/</guid><pubDate>Mon, 01 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h3 id=&quot;딥러닝의-key-component&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%94%A5%EB%9F%AC%EB%8B%9D%EC%9D%98-key-component&quot; aria-label=&quot;딥러닝의 key component permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;딥러닝의 Key Component&lt;/h3&gt;&lt;ol&gt;&lt;li&gt;The &lt;strong&gt;&lt;code&gt;data&lt;/code&gt;&lt;/strong&gt; that the model can learn from&lt;ul&gt;&lt;li&gt;다루고자 하는 데이터&lt;/li&gt;&lt;li&gt;풀고자 하는 문제의 타입에 따라 달라진다.&lt;ul&gt;&lt;li&gt;문제 타입 : Classification, Sementic Segmentation, Detection, Pose Estimation, Visual QnA 등&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;The &lt;strong&gt;&lt;code&gt;model&lt;/code&gt;&lt;/strong&gt; how to transform the data&lt;ul&gt;&lt;li&gt;데이터를 변형하여 원하는 결과를 도출하는 모델&lt;/li&gt;&lt;li&gt;AlexNet, GoogLeNet, ResNet, DenseNet, LSTM, Deep Auto Encoders, GAN 등&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;The loss function that quantifies the badness of the model&lt;ul&gt;&lt;li&gt;모델을 학습시키기 위한 손실함수&lt;/li&gt;&lt;li&gt;상황에 따라 사용하는 손실함수가 다르다.&lt;ul&gt;&lt;li&gt;Regression Task : MSE(Mean Squared Error) - [오차제곱]의 평균&lt;/li&gt;&lt;li&gt;Classification Task : CE(Cross Entropy) - [정답레이블*로그 추정치]의 평균&lt;/li&gt;&lt;li&gt;Probabilistic Task : MLE(Maximum likelihood Estimation) 또는 MSE&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;그러나 고정된 것은 아니며, 각 손실함수의 특징과 문제의 특징을 잘 고려하여 그때그때 선택해야한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;The algorithm to adjust the parameters to minimize the loss&lt;ul&gt;&lt;li&gt;목적을 달성하기 위해 손실함수를 최소화시키는 알고리즘&lt;/li&gt;&lt;li&gt;SGD, Momentum, NAG, Adagrad, Adadelta, Rmsprop 등&lt;/li&gt;&lt;li&gt;손실함수를 단순히 최소화시키는 것이 아니라, 학습하지 않은 데이터에서 잘 동작하도록 하는 것이 중요하다.&lt;ul&gt;&lt;li&gt;Dropout, Early stopping, k-fold validation, Weight decay, Batch normalization, MixUp, Ensemble, Bayesian Optimization 등&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;위의 4가지가 딥러닝의 중요한 4요소라고 볼수 있으며, 새로운 논문을 이해할 때 저 4가지에서 어떤 차이점이 있는지를 중점적으로 보게 된다.&lt;/p&gt;&lt;h1 id=&quot;딥러닝의-역사&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%94%A5%EB%9F%AC%EB%8B%9D%EC%9D%98-%EC%97%AD%EC%82%AC&quot; aria-label=&quot;딥러닝의 역사 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;딥러닝의 역사&lt;/h1&gt;&lt;p&gt;&lt;a href=&quot;https://dennybritz.com/blog/deep-learning-most-important-ideas/&quot;&gt;Deep Learning&amp;#x27;s Most Important Ideas - A Brief Historical Review(2020-07-29, Denny Britz)&lt;/a&gt;를 참고하였습니다.&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;blockquote&gt;&lt;p&gt;2012 - AlexNet&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;2012년 처음으로 Alex-net이라는 이미지넷 대회에서 딥러닝이 우승했다. 그 이후로는 한번도 딥러닝이 아닌 다른 기법이 우승한 적이 없다.&lt;/p&gt;&lt;/blockquote&gt;&lt;p&gt;그 이후로 이론적으로만 연구되던 딥러닝이 실제로 사용되기 시작했고, 패러다임의 변화가 일어났다.&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;blockquote&gt;&lt;p&gt;2013 - DQN&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;추후에 알파고를 만들게 되는 딥마인드가, 아타리 게임을 클리어하기 위하여 &lt;a href=&quot;https://ko.wikipedia.org/wiki/Q_%EB%9F%AC%EB%8B%9D&quot;&gt;Q-Learning&lt;/a&gt;을 딥러닝에 접목한 DQN을 사용하고, 논문을 냈다. 딥마인드는 이 결과를 눈여겨본 구글에 인수되었다.&lt;/p&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;2014 - Encoder / Decoder, Adam Optimizer&lt;/p&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;a href=&quot;https://ko.wikipedia.org/wiki/%EC%8B%A0%EA%B2%BD%EB%A7%9D_%EA%B8%B0%EA%B3%84_%EB%B2%88%EC%97%AD&quot;&gt;NMT(Neural Machine Translation)&lt;/a&gt;를 풀기위해 Encoder / Decoder가 등장했다.&lt;/p&gt;&lt;/blockquote&gt;&lt;p&gt;특정 언어의 시퀀스를 Encoder가 어떤 벡터에 Encoding하고, Decoder가 이 벡터를 다른 언어 시퀀스로 Decoding 시켜준다.&lt;/p&gt;&lt;p&gt;Adam(Adaptive Momentum) Optimizer가 등장했다.&lt;/p&gt;&lt;p&gt;기존의 SGD, Adam 등의 Optimizer에 비하여 Optimizing 결과가 좋아서, 많이 사용된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;여기에 이유가 설명되어 있지는 않다.&lt;/li&gt;&lt;li&gt;컴퓨팅 리소스가 충분한 Google 등의 테크기업들이 실험결과를 제공한 것을 바탕으로, &amp;#x27;그냥 그렇게 하면 좋더라...&amp;#x27;의 느낌으로 사용한다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;blockquote&gt;&lt;p&gt;2015 - GAN, ResNet&lt;/p&gt;&lt;h1 id=&quot;-7&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-7&quot; aria-label=&quot; 7 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이안 굿펠로우의 GAN(Generative Adversarial Network) 논문이 발표되었다.&lt;/p&gt;&lt;/blockquote&gt;&lt;p&gt;Generator와 Discriminator를 두고, 서로 경쟁시킴으로써 &amp;#x27;그럴듯한&amp;#x27; 결과를 만들어내는 기술이다.&lt;/p&gt;&lt;p&gt;ResNet 논문도 발표되었다.&lt;/p&gt;&lt;p&gt;Deep Learning은 Shallow Network를 활용하지 않고, 신경망의 레이어를 많이 쌓는 방식이다(그렇다고 알려져 있다). 그러나 ResNet이 나오기 이전까지는, 레이어를 너무 깊게 쌓으면 테스트 데이터에 대한 성능이 좋지 않다고 알려져 있었다.&lt;/p&gt;&lt;p&gt;ResNet은 100단이 넘어가는 수의 레이어를 쌓아도 성능이 개선되는 것을 보여주면서, &amp;#x27;딥&amp;#x27;러닝으로의 진정한 패러다임 전환이 일어났다.&lt;/p&gt;&lt;h1 id=&quot;-8&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-8&quot; aria-label=&quot; 8 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;blockquote&gt;&lt;p&gt;2017 - Transformer&lt;/p&gt;&lt;h1 id=&quot;-9&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-9&quot; aria-label=&quot; 9 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;a href=&quot;https://arxiv.org/pdf/1706.03762.pdf&quot;&gt;Attention Is All You Need&lt;/a&gt; 논문이 발표되었다.&lt;/p&gt;&lt;/blockquote&gt;&lt;p&gt;발표 직후에는 domain specific한 방법이었지만, 기존의 CNN과 RNN 등 대부분의 방법을 대체하면서 현재는 Vision까지 넘보고 있다.&lt;/p&gt;&lt;p&gt;Transformer 구조 또는 Attention 구조는 굉장히 중요한 파트이고, 왜 기존의 방법들에 비해 높은 성능을 내는지 알아두는 것이 좋다.&lt;/p&gt;&lt;h1 id=&quot;-10&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-10&quot; aria-label=&quot; 10 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;blockquote&gt;&lt;p&gt;2018 - Bert&lt;/p&gt;&lt;h1 id=&quot;-11&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-11&quot; aria-label=&quot; 11 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;Bidirection Encoder Representations from Transformers의 약자로, 2017년 발표된 Transformer 구조를 활용하되, Bidirectional Encoder를 사용한다.&lt;/p&gt;&lt;/blockquote&gt;&lt;p&gt;Bert에서 주목해야 할 것은 Bert 그 자체보다 Fine Tuned NLP Model의 등장이다.&lt;/p&gt;&lt;p&gt;Language 모델은 이전에 단어가 주어졌을 때 다음에 어떤 단어가 나올지를 맞추는 것인데, 이를 이용해 그럴싸한 문장을 만들어 내기 위해서는 해당 도메인과 관련된 수많은 말뭉치가 필요했다. 그러나 Fine Tuned NLP Model은 먼저 세상에 널려있는 수많은 말들로 Pre-tuning을 하고, 그 이후 해당 도메인의 말뭉치로 Fine-tuning을 한다.&lt;/p&gt;&lt;h1 id=&quot;-12&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-12&quot; aria-label=&quot; 12 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;blockquote&gt;&lt;p&gt;2019 - Big Language Models(GPT-X)&lt;/p&gt;&lt;h1 id=&quot;-13&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-13&quot; aria-label=&quot; 13 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;OpenAI에서 발표한 GPT-3 모델은 Bert의 끝판왕이라고 볼 수 있다. 쉽사리 학습할 수 없는 굉장히 많은(175B) 파라미터를 가지고 있다.&lt;/p&gt;&lt;h1 id=&quot;-14&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-14&quot; aria-label=&quot; 14 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;2020 - Self Supervised Learning&lt;/p&gt;&lt;h1 id=&quot;-15&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-15&quot; aria-label=&quot; 15 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;a href=&quot;https://arxiv.org/pdf/2002.05709.pdf&quot;&gt;SimCLR : a simple framework for contrastive learning of visual representations&lt;/a&gt;&lt;/p&gt;&lt;/blockquote&gt;&lt;p&gt;한정된 학습데이터가 주어졌을 때, 여러 번 수정을 해가며 가장 결과가 좋은 모델을 수동적으로 찾아내는것이 일반적인 방법이었다.&lt;/p&gt;&lt;p&gt;반면, Self Supervised Learning은 Label을 모르는 unsupervised data를 활용할 수 있다.&lt;/p&gt;&lt;p&gt;SimCLR 논문에서는 어떻게 좋은 Visual Representation(이미지를 벡터로 바꾸는것)을 할 수 있을 것인가를 다룬다.&lt;/p&gt;&lt;p&gt;최근에는 이 논문을 기반으로 &lt;a href=&quot;https://arxiv.org/abs/2006.07733&quot;&gt;BYOL(Bootstrap Your Own Latent)&lt;/a&gt;와 같은 핫한 논문들이 나오고 있다.&lt;/p&gt;&lt;p&gt;또, 고도화된 도메인 지식이 있을 때 학습데이터를 시뮬레이터로 추가로 만들어내는, self supervised data sampling과 같은 기법도 등장하고 있다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[베이즈 정리와 인과관계]]></title><description><![CDATA[베이즈 통계학 맛보기 by 임성빈 교수님, BoostCamp AI Tech 3주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/11_bayes/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/11_bayes/</guid><pubDate>Mon, 01 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;베이즈-정리&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%B2%A0%EC%9D%B4%EC%A6%88-%EC%A0%95%EB%A6%AC&quot; aria-label=&quot;베이즈 정리 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;베이즈 정리&lt;/h1&gt;&lt;p&gt;이전에 배웠던 통계학은 &lt;strong&gt;&lt;code&gt;빈도주의(frequent statistics)&lt;/code&gt;&lt;/strong&gt;에 해당하는 내용이었다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;빈도주의란, 우리가 익숙한 확률에 대한 전통적인 관점이다.&lt;/li&gt;&lt;li&gt;그 사건이 일어날 횟수의 장기적인 비율(경향)을 연역적으로 정의한 것이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이번에는 딥러닝과 좀 더 직접적인 연관이 있는 &lt;strong&gt;&lt;code&gt;베이즈 정리&lt;/code&gt;&lt;/strong&gt;에 대해 알아보자. &lt;strong&gt;&lt;code&gt;베이즈 정리&lt;/code&gt;&lt;/strong&gt;는 &lt;strong&gt;모델의 모수를 추정할 때 사용하는 기법&lt;/strong&gt;이다. &lt;strong&gt;데이터가 새로 추가될 때 정보를 업데이트하는 이론적 방법&lt;/strong&gt;이기도 하다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;베이지안은, 확률에 대한 새로운 관점이다.&lt;/li&gt;&lt;li&gt;빈도주의와 달리 확률을 &lt;strong&gt;&amp;#x27;주장에 대한 신뢰도&amp;#x27;&lt;/strong&gt;로 판단한다.&lt;/li&gt;&lt;li&gt;이 과정에서, 반복의 개념은 사용되지 않는다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;먼저-알아-둘-것---조건부-확률&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%A8%BC%EC%A0%80-%EC%95%8C%EC%95%84-%EB%91%98-%EA%B2%83---%EC%A1%B0%EA%B1%B4%EB%B6%80-%ED%99%95%EB%A5%A0&quot; aria-label=&quot;먼저 알아 둘 것   조건부 확률 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;먼저 알아 둘 것 - 조건부 확률&lt;/h3&gt;&lt;p&gt;베이즈 정리는 조건부확률을 이용하여 정보를 갱신하는 방법을 알려준다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mo&gt;∩&lt;/mo&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(A\cap B) = P(B)P(A|B)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∩&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;조건부확률 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(A|B)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 사건 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;B&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 가 일어난 상황에서 사건 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;A&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 가 발생할 확률을 의미한다.&lt;/p&gt;&lt;h2 id=&quot;베이즈-정리란&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%B2%A0%EC%9D%B4%EC%A6%88-%EC%A0%95%EB%A6%AC%EB%9E%80&quot; aria-label=&quot;베이즈 정리란 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;베이즈 정리란?&lt;/h2&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mo&gt;∩&lt;/mo&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(B|A) = \frac{P(A \cap B)}{P(A)} = P(B)\frac{P(A|B)}{P(A)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.363em;vertical-align:-0.936em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;∩&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.363em;vertical-align:-0.936em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;A가 조건부로 주어졌을 때 B가 일어날 확률 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(B|A)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 B가 조건부로 주어졌을 때 A가 일어날 확률 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(A|B)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(B)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 곱하고, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(A)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 나누는것과 같다.&lt;/p&gt;&lt;p&gt;이 때 위의 식을 잘 살펴보면, A라는 새로운 정보가 주어졌을 때, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(B)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로부터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;B&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(B|A)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05017em&quot;&gt;B&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를  계산하는 방법을 알 수 있다. 즉, &lt;strong&gt;새로운 정보를 토대로 어떤 사건이 발생했다는 주장에 대한 신뢰도를 갱신&lt;/strong&gt;해 나갈 수 있을 것이다. 이것이 &lt;strong&gt;&lt;code&gt;베이즈 정리&lt;/code&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\theta|\mathcal{D}) = P(\theta)\frac{P(\mathcal{D}|\theta)}{P(\mathcal{D})}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.363em;vertical-align:-0.936em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;좌변&lt;ul&gt;&lt;li&gt;여기서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathcal{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;strong&gt;새로 관찰하는 데이터(즉, 새로운 정보)&lt;/strong&gt;를 의미한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;strong&gt;Hypothesis&lt;/strong&gt;, 또는 &lt;strong&gt;모델링하는 이벤트&lt;/strong&gt;, 또는 모델에서 계산하고 싶어하는 &lt;strong&gt;모수(parameter)&lt;/strong&gt;를 의미한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\theta|\mathcal{D})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;strong&gt;&lt;code&gt;사후확률(posterior probability)&lt;/code&gt;&lt;/strong&gt;로, &lt;strong&gt;데이터가 주어져 있을 때, hypothesis가 성립할 확률&lt;/strong&gt;을 의미한다. 즉, &lt;strong&gt;갱신하고 난 후의 신뢰도&lt;/strong&gt;를 의미한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;우변&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\theta)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;strong&gt;&lt;code&gt;사전확률(prior distribution)&lt;/code&gt;&lt;/strong&gt;로, 모델링하고자 하는 타겟(모수 등)에 대해 &lt;strong&gt;데이터를 분석하기 전에 가정한 확률분포&lt;/strong&gt;를 의미한다. 즉, &lt;strong&gt;갱신하기 전의 신뢰도&lt;/strong&gt;를 의미한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\mathcal{D}|\theta)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;strong&gt;&lt;code&gt;가능도(likelihood)&lt;/code&gt;&lt;/strong&gt;로, &lt;strong&gt;현재 주어진 [모수/가정]에서 데이터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathcal{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 관찰될 확률&lt;/strong&gt;을 의미한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\mathcal{D})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;strong&gt;&lt;code&gt;Evidence&lt;/code&gt;&lt;/strong&gt;를 의미한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 때 , 베이즈 정리를 적용하여 현재의 가정(신뢰도)에서 새로운 데이터에 대한 신뢰도를 얻어낼 수 있다.&lt;/p&gt;&lt;h3 id=&quot;베이즈-정리-예제---코로나-바이러스-99&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%B2%A0%EC%9D%B4%EC%A6%88-%EC%A0%95%EB%A6%AC-%EC%98%88%EC%A0%9C---%EC%BD%94%EB%A1%9C%EB%82%98-%EB%B0%94%EC%9D%B4%EB%9F%AC%EC%8A%A4-99&quot; aria-label=&quot;베이즈 정리 예제   코로나 바이러스 99 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;베이즈 정리 예제 - 코로나 바이러스 99&lt;/h3&gt;&lt;p&gt;좀 더 이해하기 쉽게 코로나 바이러스의 예제를 들어보자.&lt;/p&gt;&lt;p&gt;COVID-99의 &lt;strong&gt;발병률이 10%&lt;/strong&gt;로 알려져 있다. COVID-99에 &lt;strong&gt;실제로 걸렸을 때 검진될 확률은 99%, 실제로 걸리지 않았을 때 오검진될 확률이 1%&lt;/strong&gt;라고 하자.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;어떤 사람이 질병에 걸렸다고 검진 결과가 나왔을 때 정말 COVID-99에 감염되었을 확률&lt;/strong&gt;은?&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;&lt;code&gt;사전확률&lt;/code&gt;, &lt;code&gt;민감도(재현율,Recall)&lt;/code&gt;, &lt;code&gt;오탐율(False alarm)&lt;/code&gt;을 가지고 &lt;code&gt;정밀도(Precision)&lt;/code&gt;을 계산하는 문제이다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;기존에 알려져있는 발병률은 사전확률 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\theta)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 볼 수 있으며, 사전에 정의된 것이므로 관찰 불가하다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;검진될 확률과 오검진된 확률을 각각 가능도 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;0.99&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\mathcal{D}|\theta) = 0.99&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;9&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;9&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;¬&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;0.01&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\mathcal{D}|\neg\theta) = 0.01&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;¬&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 볼 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathcal{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 테스트 결과라고 정의하므로 관찰 가능하다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;여기서 베이즈 정리에 필요한 식에 빠진 것은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\mathcal{D})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, Evidence다. Evidence는 어떻게 계산할 수 있을까?&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/munder&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;0.99&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;0.1&lt;/mn&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;0.001&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;0.9&lt;/mn&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;0.108&lt;/mn&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
P(\mathcal{D})=\sum_\theta P(\mathcal{D}|\theta)P(\theta) &amp;amp;= 0.99 \times 0.1 + 0.001 \times 0.9 \\
&amp;amp;= 0.108
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:4.152118000000001em;vertical-align:-1.8260589999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.3260590000000008em&quot;&gt;&lt;span style=&quot;top:-4.326059000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000005em&quot;&gt;&lt;span style=&quot;top:-1.847887em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3021129999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.8839460000000008em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.8260589999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.3260590000000008em&quot;&gt;&lt;span style=&quot;top:-4.326059000000001em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;9&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;9&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;9&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-1.8839460000000008em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.0500050000000005em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;8&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.8260589999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;가능도를 가지고 계산할 수 있다. &lt;code&gt;주변분포(marginal distribution)&lt;/code&gt;를 계산하는 방법과 &lt;strong&gt;조건부확률의 정의&lt;/strong&gt;를 이용하면 된다.&lt;/li&gt;&lt;li&gt;쉽게 풀이하면, &lt;div&gt;[실제로 감염되었을 확률 *  검진될 확률] + [실제로 감염되지 않았을 확률 * 오검진될 확률]&lt;/div&gt;이라고 볼 수 있다.&lt;ul&gt;&lt;li&gt;만약 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;¬&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\mathcal{D}|\neg\theta)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;¬&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 모르면 이 문제는 풀기 어렵다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;위의 식들을 종합해, 갱신된 신뢰도, 즉 문제의 정답은 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;0.1&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;0.99&lt;/mn&gt;&lt;mn&gt;0.108&lt;/mn&gt;&lt;/mfrac&gt;&lt;mo&gt;≈&lt;/mo&gt;&lt;mn&gt;0.916&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\theta|\mathcal{D}) = 0.1 \times \frac{0.99}{0.108} \approx 0.916&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.00744em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;8&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;9&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;9&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;9&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;6&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;이 때, &lt;strong&gt;오탐율(False alarm)이 오르면 테스트 정밀도(Precision)가 당연히 떨어진다.&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;예를 들어 오탐율 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;¬&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\mathcal{D}|\neg\theta)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;¬&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 0.01이 아닌 0.1이라면, 정밀도는 0.916이 아닌 0.524가 된다.&lt;ul&gt;&lt;li&gt;이처럼, &lt;strong&gt;아무리 민감도가 높더라도 오탐율이 높아지면 정밀도가 크게 떨어진다.&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;조건부-확률의-시각화&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A1%B0%EA%B1%B4%EB%B6%80-%ED%99%95%EB%A5%A0%EC%9D%98-%EC%8B%9C%EA%B0%81%ED%99%94&quot; aria-label=&quot;조건부 확률의 시각화 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;조건부 확률의 시각화&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:49.609375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;conditional_prob_visualization&quot; title=&quot;conditional_prob_visualization&quot; src=&quot;/static/789d54b180f168f80477e6181ef6cfc1/2bef9/conditional_prob_visualization.png&quot; srcSet=&quot;/static/789d54b180f168f80477e6181ef6cfc1/6f3f2/conditional_prob_visualization.png 256w,/static/789d54b180f168f80477e6181ef6cfc1/01e7c/conditional_prob_visualization.png 512w,/static/789d54b180f168f80477e6181ef6cfc1/2bef9/conditional_prob_visualization.png 1024w,/static/789d54b180f168f80477e6181ef6cfc1/71c1d/conditional_prob_visualization.png 1536w,/static/789d54b180f168f80477e6181ef6cfc1/6297e/conditional_prob_visualization.png 2044w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;[일치여부(신뢰성) + 판단결과] 형태로 표시한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;음성이 나왔을 때 양성으로 판단&lt;/strong&gt;하는 것을 &lt;strong&gt;&lt;code&gt;False Positive&lt;/code&gt;&lt;/strong&gt;라고 하고, &lt;code&gt;1종 오류&lt;/code&gt;로 친다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;양성이 나왔을 때 음성으로 판단&lt;/strong&gt;하는것을 &lt;strong&gt;&lt;code&gt;False Negative&lt;/code&gt;&lt;/strong&gt;라고 하고 &lt;code&gt;2종 오류&lt;/code&gt;로 친다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;데이터 분석의 성격에 따라 1종 오류와 2종 오류 중 어느 것을 우선하여 줄일 지가 달라진다.&lt;/p&gt;&lt;p&gt;특히 2종 오류(False Negative)같은 경우 의료 분야에서 아주 심각한 문제이다. 누군가가 큰 병에 걸렸는데 이를 오진할 경우, 목숨까지 위험할 수 있다.&lt;/p&gt;&lt;p&gt;이에 비해 1종 오류(False Negative)는 비교적 위험성이 떨어지므로, 보통은 오탐율(False Alarm)을 희생하더라도, 민감도(Recall)을 최대한 줄이는 방식으로 설계하곤 한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:45.3125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;precision&quot; title=&quot;precision&quot; src=&quot;/static/07c05a576da6bda5196d02d2f1fd9f22/2bef9/precision.png&quot; srcSet=&quot;/static/07c05a576da6bda5196d02d2f1fd9f22/6f3f2/precision.png 256w,/static/07c05a576da6bda5196d02d2f1fd9f22/01e7c/precision.png 512w,/static/07c05a576da6bda5196d02d2f1fd9f22/2bef9/precision.png 1024w,/static/07c05a576da6bda5196d02d2f1fd9f22/71c1d/precision.png 1536w,/static/07c05a576da6bda5196d02d2f1fd9f22/a878e/precision.png 2048w,/static/07c05a576da6bda5196d02d2f1fd9f22/b5c21/precision.png 2154w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;정밀도를 계산할 때에는, True Positive와 False Negative를 더한 것, 즉 양성판정률을 분모로 삼는다. 따라서 오탐율이 낮을수록 분모가 커지므로 정밀도가 높아지고, 민감도가 커질수록 분자가 커지므로 정밀도가 높아진다.&lt;/p&gt;&lt;p&gt;이처럼 &lt;strong&gt;정밀도는 오탐율과 민감도에 영향을 받는다.&lt;/strong&gt;&lt;/p&gt;&lt;p&gt;위의 코로나 바이러스 99 예제에서, 앞서서 COVID-99 양성 판정을 받은 사람이 두 번째 검진에도 양성이 나왔을 때, 진짜 COVID-99에 걸렸을 확률은 얼마나 될까?&lt;/p&gt;&lt;p&gt;이전에 계산한 사후확률을 사전확률로 활용하여, 베이즈 정리로 다시 한번 정보를 업데이트해보자.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mtable rowspacing=&quot;0.24999999999999992em&quot; columnalign=&quot;right left&quot; columnspacing=&quot;0em&quot;&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;0.99&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;0.524&lt;/mn&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mn&gt;0.1&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mn&gt;0.476&lt;/mn&gt;&lt;mo&gt;≈&lt;/mo&gt;&lt;mn&gt;0.566&lt;/mn&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;mtr&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;mo&gt;∗&lt;/mo&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;mtd&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;0.524&lt;/mn&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;0.99&lt;/mn&gt;&lt;mn&gt;0.566&lt;/mn&gt;&lt;/mfrac&gt;&lt;mo&gt;≈&lt;/mo&gt;&lt;mn&gt;0.917&lt;/mn&gt;&lt;/mrow&gt;&lt;/mstyle&gt;&lt;/mtd&gt;&lt;/mtr&gt;&lt;/mtable&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\begin{aligned}
P(\mathcal{D}^*) &amp;amp;= 0.99 \times 0.524 + 0.1 \times 0.476 \approx 0.566 \\
P(\theta|\mathcal{D}^*) &amp;amp;= 0.524 \times \frac{0.99}{0.566} \approx 0.917
\end{aligned}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.8074399999999997em;vertical-align:-1.6537199999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mtable&quot;&gt;&lt;span class=&quot;col-align-r&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.15372em&quot;&gt;&lt;span style=&quot;top:-4.63516em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.32144em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.738696em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mbin mtight&quot;&gt;∗&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.6537200000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.32144em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.738696em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mbin mtight&quot;&gt;∗&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6537199999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;col-align-l&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:2.15372em&quot;&gt;&lt;span style=&quot;top:-4.63516em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.32144em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;9&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;9&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;4&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;4&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;7&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;6&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;6&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;6&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.6537200000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.32144em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;4&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;6&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;6&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;9&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;9&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;9&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;7&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6537199999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이처럼 한번의 검진율로는 0.524밖에 되지 않았던 신뢰성이, 한번 더 정보를 업데이트함으로써 0.917까지 높아지게 되었다. 세번째 검사까지 하면 0.991까지 업데이트된다.&lt;/p&gt;&lt;p&gt;이처럼 베이즈 정리는 &lt;strong&gt;데이터를 추가할 때마다 업데이트된 신뢰성을 얻을 수 있으므로 아주 유효한 방식&lt;/strong&gt;이다.&lt;/p&gt;&lt;h2 id=&quot;조건부확률-→-인과관계&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A1%B0%EA%B1%B4%EB%B6%80%ED%99%95%EB%A5%A0-%E2%86%92-%EC%9D%B8%EA%B3%BC%EA%B4%80%EA%B3%84&quot; aria-label=&quot;조건부확률 → 인과관계 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;조건부확률 → 인과관계?&lt;/h2&gt;&lt;p&gt;조건부 확률을 사용할 때 범하기 쉬운 오류가 있다. &lt;strong&gt;조건부 확률이 &lt;code&gt;인과관계&lt;/code&gt;로 이어진다고 생각하는 오류&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;조건부 확률은 유용한 통계적 해석을 제공하지만, 인과관계(Causality)를 추론할 때 함부로 사용해서는 안된다. &lt;strong&gt;&lt;div&gt;데이터가 아무리 많아져도 조건부 확률만을 가지고 인과관계를 추론하는 것은 불가능&lt;/div&gt;&lt;/strong&gt;하다.&lt;/p&gt;&lt;h3 id=&quot;인과관계&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9D%B8%EA%B3%BC%EA%B4%80%EA%B3%84&quot; aria-label=&quot;인과관계 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;인과관계&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;인과관계&lt;/code&gt;&lt;/strong&gt;는 &lt;strong&gt;데이터 분포의 변화에 강건&lt;/strong&gt;한 예측모형을 만들 때 필요하다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:31.640625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA1UlEQVQY02VQ7Y4DIQjs+z9os9e6KgiooO7hNv1xuQkSHRm+HtdfrNv/nPAMiZkBgIgQCwtf//DwM+Ya63KzsfzedCbg5yuGmKEQECcoSGJzVR1QxFORNI/c4mYjnDHnXIhBDLmnGOMZkZiqpowpZyJJ1EE0pRjCCUhqX7HU5kq1wW2bVny9g5eVbrWpSDUbzabXs1pCCNnF4xavdc21mwdkT5kyeZ/StOt03u6JCtc9QqEIpeno5l/rgYjHcYR3MLPl2OSGP9XRt9t7/LBresBnW87/AujCXHY3tPZlAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;causality&quot; title=&quot;causality&quot; src=&quot;/static/1344d658a15ce68b8a5b1f73ff2260a6/2bef9/causality.png&quot; srcSet=&quot;/static/1344d658a15ce68b8a5b1f73ff2260a6/6f3f2/causality.png 256w,/static/1344d658a15ce68b8a5b1f73ff2260a6/01e7c/causality.png 512w,/static/1344d658a15ce68b8a5b1f73ff2260a6/2bef9/causality.png 1024w,/static/1344d658a15ce68b8a5b1f73ff2260a6/71c1d/causality.png 1536w,/static/1344d658a15ce68b8a5b1f73ff2260a6/a878e/causality.png 2048w,/static/1344d658a15ce68b8a5b1f73ff2260a6/922e6/causality.png 2140w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;데이터 분포의 상황이 바뀌는 경우가 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;새로운 정책 / 치료법등을 도입했을 때&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 경우 조건부 확률만을 가지고 예측 모형을 만들었을 때, &lt;strong&gt;시나리오에 따라 예측정확도의 변동성이 크다&lt;/strong&gt;. 이에 비해 인과관계 기반 예측모형은 &lt;strong&gt;시나리오가 변한다고 하더라도 정확도가 크게 변하지 않는다&lt;/strong&gt;. 다만, 인과관계만 고려해서는 높은 예측 정확도를 담보하기 어렵다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:19.140625000000004%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAECAIAAAABPYjBAAAACXBIWXMAABYlAAAWJQFJUiTwAAAAbElEQVQI132OSQ4DIQwE+f8rcxtFMMJmM4shaSXKYSQmdbB86GrbvG5orTnnsDxPOjmWkkFKKf9AwPyRi8hUpZAfh2Um7z0zxxjpg4hs5N47Eqo656y1fvftgYu81sLEbyEEmGgZY1hrUbGV35Oe6pySasIFAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;confounding_factor&quot; title=&quot;confounding_factor&quot; src=&quot;/static/18bb6b0cc2a81d0bbb2b12da4f59ac29/2bef9/confounding_factor.png&quot; srcSet=&quot;/static/18bb6b0cc2a81d0bbb2b12da4f59ac29/6f3f2/confounding_factor.png 256w,/static/18bb6b0cc2a81d0bbb2b12da4f59ac29/01e7c/confounding_factor.png 512w,/static/18bb6b0cc2a81d0bbb2b12da4f59ac29/2bef9/confounding_factor.png 1024w,/static/18bb6b0cc2a81d0bbb2b12da4f59ac29/71c1d/confounding_factor.png 1536w,/static/18bb6b0cc2a81d0bbb2b12da4f59ac29/a878e/confounding_factor.png 2048w,/static/18bb6b0cc2a81d0bbb2b12da4f59ac29/55681/confounding_factor.png 2062w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;인과관계를 알아내기 위해서는, &lt;strong&gt;&lt;code&gt;중첩요인(confounding factor)&lt;/code&gt;&lt;/strong&gt; &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;Z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 효과를 제거하고 원인에 해당하는 변수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;T&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;T&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;T&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;만의 인과관계를 계산해야 한다. &lt;/p&gt;&lt;p&gt;간단한 예로, 키와 지능의 연관성을 들어볼 수 있다.&lt;/p&gt;&lt;p&gt;일반적으로 키와 지능은 상관관계가 없는 특징이라고 여겨진다. 그러나 전체인구통계에서, 키와 지능을 살펴보면 이상하게도 키가 클수록 지능이 높다는 것을 알 수 있다.&lt;/p&gt;&lt;p&gt;그렇다면 과연 키와 지능은 인과관계가 있는 것일까? 키가 클수록 지능이 높아지게 되는걸까?&lt;/p&gt;&lt;ul&gt;&lt;li&gt;그렇지 않다. 이는 &amp;#x27;나이&amp;#x27;라는 중첩요인 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;Z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 제거하지 않았기 때문이다.&lt;/li&gt;&lt;li&gt;성장기에서, 나이가 많아지면 키도 커지고 지능도 높아진다.&lt;/li&gt;&lt;li&gt;키 → 지능으로 바로 연결되는 것이 아니라, 모든 조사대상의 나이를 동일하게 맞추어 나이라는 중첩요인을 제거한 뒤 조사해야한다.&lt;ul&gt;&lt;li&gt;그러지 않으면 &lt;code&gt;가짜 연관성(spurious correlation)&lt;/code&gt;이 나온다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;인과관계-추론-예제--치료법에-따른-완치율&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9D%B8%EA%B3%BC%EA%B4%80%EA%B3%84-%EC%B6%94%EB%A1%A0-%EC%98%88%EC%A0%9C--%EC%B9%98%EB%A3%8C%EB%B2%95%EC%97%90-%EB%94%B0%EB%A5%B8-%EC%99%84%EC%B9%98%EC%9C%A8&quot; aria-label=&quot;인과관계 추론 예제  치료법에 따른 완치율 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;인과관계 추론 예제 : 치료법에 따른 완치율&lt;/h3&gt;&lt;p&gt;신장결석에 대한 치료법이 두 개가 있다. 두 치료법을 따랐을 때, 신장결석 크기에 따른 완치율 통계 결과가 다음과 같이 나와있다면, 어떤 치료법을 고르는 것이 더 나을까?&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:32.421875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAGCAIAAABM9SnKAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA5UlEQVQY01VQWY6EIBT0/mfr9AF63DvtwioKAorQpU4mmfowL+/VJllKMaWklOJcpAsxxn3fQwje+23btNZKzc77oiirqgYzz4uyrHDNbrZzzlp3HMetF0L2/WCthZgQ+vl0sICyaVoY1XXTtm9cM8hwuKIOYwxmiJHxeDyXReNKL0yTklIyxkGANaXsTBZCaLMiEz1hNgwjxF0/vH7y1RgwKBRcIBBfuMzzjMU4klPMGFut/ftb9MQAI8yET1SqEHbwUMGYFVj0WQcblM2QCV76j3sj1FJ0xDoHAYR4Pnchxt+n+QIyUldLDqyxoAAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;causality_ex&quot; title=&quot;causality_ex&quot; src=&quot;/static/0560c8317a34b5a67b85bb265b787e9c/2bef9/causality_ex.png&quot; srcSet=&quot;/static/0560c8317a34b5a67b85bb265b787e9c/6f3f2/causality_ex.png 256w,/static/0560c8317a34b5a67b85bb265b787e9c/01e7c/causality_ex.png 512w,/static/0560c8317a34b5a67b85bb265b787e9c/2bef9/causality_ex.png 1024w,/static/0560c8317a34b5a67b85bb265b787e9c/71c1d/causality_ex.png 1536w,/static/0560c8317a34b5a67b85bb265b787e9c/a878e/causality_ex.png 2048w,/static/0560c8317a34b5a67b85bb265b787e9c/9e7e4/causality_ex.png 2056w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;전체적인 치료법의 성공률(overall)을 보면 b가 더 나아 보인다.&lt;/p&gt;&lt;p&gt;그러나, 실제로는 신장결석 여부와 관계없이 각각의 환자군에서 a가 모두 성공률이 더 높다. 어떻게 이런 결과가 나오게 된 것일까?&lt;/p&gt;&lt;p&gt;이는 &lt;a href=&quot;https://bioinformaticsandme.tistory.com/117&quot;&gt;심슨의 역설&lt;/a&gt;이라고 불리는 아주 유명한 패러독스로, 부분이 크다고해서 전체도 큰것은 아니라는 것을 나타낸다. 이를 명확히 꿰뚫어보기 위해서는 전체를 부분별로 각각 나누어 보아, 정확한 원인분석을 해야한다.&lt;/p&gt;&lt;p&gt;위의 예제로 따지자면, 단순히 치료법에 따른 조건부확률만 계산해서는 안되고, 신장 결석 크기에 따른 중첩효과를 제거해야한다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:16.015625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAADCAIAAAAcOLh5AAAACXBIWXMAABYlAAAWJQFJUiTwAAAAj0lEQVQI1yWN2w7CIBBE+/9f5qvPNhKDCCXltizIzeJW92EymcnZWY45q3weOc05xxhm3xGx1qqUstaGEHLOCbGRpmSMIW29z98tL+1CLhRhjJiScw4ACOaca63JU3K2AH/vva+ltHf5tL64gDbECECDkXjEUs5fQghSxti2bQTcGaPqtq5aKngIebnmlX8BOO+ph11DW8gAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;intervention&quot; title=&quot;intervention&quot; src=&quot;/static/1162593b83476c53ca3f9c58edf0eb3e/2bef9/intervention.png&quot; srcSet=&quot;/static/1162593b83476c53ca3f9c58edf0eb3e/6f3f2/intervention.png 256w,/static/1162593b83476c53ca3f9c58edf0eb3e/01e7c/intervention.png 512w,/static/1162593b83476c53ca3f9c58edf0eb3e/2bef9/intervention.png 1024w,/static/1162593b83476c53ca3f9c58edf0eb3e/71c1d/intervention.png 1536w,/static/1162593b83476c53ca3f9c58edf0eb3e/a878e/intervention.png 2048w,/static/1162593b83476c53ca3f9c58edf0eb3e/ade5e/intervention.png 2188w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:16.796875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAADCAIAAAAcOLh5AAAACXBIWXMAABYlAAAWJQFJUiTwAAAAk0lEQVQI1x2NzQ7CMAyD9/6PxpET7LC10KENsf4lTdtM7Qj4Esv6Yg89U14e/fyp1uqdQwBm3rbNWuu9J6IEUBBTIkkQ8WD+4+dg3hb5gBhDjACw77vQrbV5np1zKckjUEq1FECpFTAelTmXxjx8XLQhZiKpzlku9d5leZommTXGKKXEaK1DCPdxXJ4G9Gu9XPGmvkx1qbhGFHX5AAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;intervention_2&quot; title=&quot;intervention_2&quot; src=&quot;/static/ec4ee78c98d2da5c1fe7a63952eac046/2bef9/intervention_2.png&quot; srcSet=&quot;/static/ec4ee78c98d2da5c1fe7a63952eac046/6f3f2/intervention_2.png 256w,/static/ec4ee78c98d2da5c1fe7a63952eac046/01e7c/intervention_2.png 512w,/static/ec4ee78c98d2da5c1fe7a63952eac046/2bef9/intervention_2.png 1024w,/static/ec4ee78c98d2da5c1fe7a63952eac046/71c1d/intervention_2.png 1536w,/static/ec4ee78c98d2da5c1fe7a63952eac046/a878e/intervention_2.png 2048w,/static/ec4ee78c98d2da5c1fe7a63952eac046/a628a/intervention_2.png 2170w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;조정(intervention)&lt;/code&gt;&lt;/strong&gt;과정을 통해 중첩효과 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;Z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 개입을 제거하여, 신장결석 크기와 관계없이 치료법 a와 b를 선택했을때의 완치율을 알 수 있다.  이는 조건부 확률을 사용하는 베이즈 정리와는 달리, &lt;strong&gt;인과관계로 예측&lt;/strong&gt;하는 것이다.&lt;/p&gt;&lt;p&gt;제거하고 나면, 치료법 a는 약 0.83, 치료법 b는 약 0.78로 치료법 a를 선택하는 것이 더 나은 방법이라는 것을 확인할 수 있다.&lt;/p&gt;&lt;p&gt;이처럼 &lt;strong&gt;중첩효과를 제거함으로써 데이터 분석 시에 좀 더 안정적인 정책 분석이나 예측모형의 설계가 가능&lt;/strong&gt;하다.&lt;/p&gt;&lt;p&gt;따라서 단순히 조건부확률만으로 데이터 분석을 하는 것은 위험하고, &lt;strong&gt;데이터에서 추론할 수 있는 사실 관계들이나 변수들끼리의 관계, 도메인 지식 등을 활용함&lt;/strong&gt;으로써 강건한 데이터모형을 만들 수 있다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[딥러닝 베이직 - 02.신경망과 다층퍼셉트론]]></title><description><![CDATA[Neural Network & Multi Layer Perceptron by 최성준 교수님, BoostCamp AI Tech 3주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/13_nn&amp;mlp/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/13_nn&amp;mlp/</guid><pubDate>Mon, 01 Feb 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 학습한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;neural-networks--multi-layer-perceptron&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#neural-networks--multi-layer-perceptron&quot; aria-label=&quot;neural networks  multi layer perceptron permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Neural Networks &amp;amp; Multi-Layer Perceptron&lt;/h1&gt;&lt;h2 id=&quot;neural-network란&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#neural-network%EB%9E%80&quot; aria-label=&quot;neural network란 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Neural Network란&lt;/h2&gt;&lt;blockquote&gt;&lt;p&gt;&amp;quot;Neural networks are computing systems vaguely inspired by the biological neural networks that constitute animal brains.&amp;quot; -- &lt;cite&gt;wekipedia&lt;/cite&gt;&lt;/p&gt;&lt;/blockquote&gt;&lt;p&gt;위키피디아의 정의는 인공신경망이 동물의 뇌신경망을 모방했다고 말한다. 그러나 최초에는 그렇게 시작했을지 몰라도, 요새 딥러닝에 사용되는 인공신경망은 다른 형태를 가진다.&lt;/p&gt;&lt;p&gt;인공신경망을 수학적으로 정의해보면 다음과 같다.&lt;/p&gt;&lt;blockquote&gt;&lt;p&gt;Neural Networks are function approximators that stack affine transformations followed by nonlinear transformations.&lt;/p&gt;&lt;/blockquote&gt;&lt;p&gt;&amp;quot;목적하는 결과를 도출해주는 어떤 함수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 근사하기 위해, 행렬-비선형 연산이 반복적으로 일어나는 함수&amp;quot;&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;linear-neural-network&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#linear-neural-network&quot; aria-label=&quot;linear neural network permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Linear Neural Network&lt;/h2&gt;&lt;ul&gt;&lt;li&gt;Data &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;D&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개의  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x,y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 순서쌍&lt;/li&gt;&lt;li&gt;Model : 추정치 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\hat{y}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.19444em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 를 도출해내는 선형식(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;wx + b)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.66666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;Loss : 추정치와 정답의 오차제곱의 평균 (MSE)&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;우리의 목적은 Loss를 최소화하는 것이므로, &lt;code&gt;Loss Function&lt;/code&gt;을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;w&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해 편미분하여, Loss가 작아지는 방향으로 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;w&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 업데이트해준다. &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;b&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해서도 마찬가지.&lt;/p&gt;&lt;p&gt;이 방식을 &lt;code&gt;Gradient Descent&lt;/code&gt;라고 하고, 만약 Loss Function이 아니라 &lt;code&gt;Reward Function&lt;/code&gt;이어서 어떤 값을 키우고싶다면 &lt;code&gt;Gradient Ascent&lt;/code&gt;를 사용한다.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;그런데, Layer는 한 층만 있는게 아니라 여러 층이 있다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;마지막 층에서 나왔던 Loss Function 값을 전체 파라미터로 다 미분을 하는 것&lt;/strong&gt;이 &lt;strong&gt;&lt;code&gt;역전파&lt;/code&gt;&lt;/strong&gt;이고, &lt;strong&gt;역전파로 나오는 각 파라미터의 편미분을 업데이트 시키는 것&lt;/strong&gt;이 실제 딥러닝에서의 &lt;strong&gt;&lt;code&gt;경사 하강법&lt;/code&gt;&lt;/strong&gt;이 된다.&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이 때, 업데이트의 &lt;code&gt;Step size(learning rate)&lt;/code&gt;가 너무 크면, 학습 중에 점프하는 크기가 너무 커져 minimum 값에 접근할 수 없고, 결국 학습이 아예 되지 않는다. 따라서 적절한 step size를 잡는 것이 중요한데, 이 때문에 &lt;code&gt;adaptive learning rate&lt;/code&gt; 같은 방법이 생겨났다.&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;multi-dimensional-inputoutput&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#multi-dimensional-inputoutput&quot; aria-label=&quot;multi dimensional inputoutput permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;multi dimensional input/output&lt;/h3&gt;&lt;p&gt;물론 다차원의 input과 output도 다룰 수 있다. 행렬을 사용하면 된다. &lt;strong&gt;&lt;div&gt;행렬 곱을 통해 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;m&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 차원에서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;차원으로 이동시킬 수 있기 때문&lt;/div&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;일차원에서는 단순한 상수였던 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;b&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 다차원에서는 각각 &lt;strong&gt;행렬&lt;/strong&gt;과 &lt;strong&gt;벡터&lt;/strong&gt;로 정의된다.&lt;/p&gt;&lt;div&gt;&lt;p&gt;어떤 행렬을 찾겠다는 것은, 서로 다른 두 차원 사이에 선형변환을 찾겠다는 의미이다. 행렬 곱을 두 개의 벡터 space 간의 변환이라고 생각하면 좋을 것이다.&lt;/p&gt;&lt;/div&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;multi-layer-perceptron&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#multi-layer-perceptron&quot; aria-label=&quot;multi layer perceptron permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Multi-Layer Perceptron&lt;/h2&gt;&lt;p&gt;선형 변환은 아무리 층을 쌓아봐야 다시 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Wx + b&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.76666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 형태로 치환된다. 따라서 인공신경망의 표현력을 극대화하기 위해서는 층마다 *&lt;strong&gt;*비선형변환 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 해주어야 한다. 그래야 &lt;/strong&gt;표현(식)의 형태가 다양해지기 때문**이다.&lt;/p&gt;&lt;p&gt;&lt;code&gt;다층퍼셉트론&lt;/code&gt;, 즉 &lt;code&gt;MLP&lt;/code&gt;는 이처럼 &lt;strong&gt;각 은닉층에 [선형변환-비선형변환]의 한 사이클을 수행&lt;/strong&gt;하는, &lt;strong&gt;여러 층의 인공신경망&lt;/strong&gt;을 일컫는다.&lt;/p&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;loss-function&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#loss-function&quot; aria-label=&quot;loss function permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Loss function&lt;/h3&gt;&lt;p&gt;&lt;code&gt;Regression Task&lt;/code&gt;에는 &lt;code&gt;MSE(Mean Squared Error)&lt;/code&gt;를 사용한다는 것을 언급했었다. 오차제곱(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L_2-loss&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)을 이용해서 추정치와 정답의 차이를 도출해낼 수 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mfrac&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/munderover&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;/munderover&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msubsup&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;MSE = \frac{1}{N}\sum^N_{i=1}\sum^D_{d=1}(y_i^{(d)}-\hat{y}_i^{(d)})^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.1304490000000005em;vertical-align:-1.302113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.8283360000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.8283360000000002em&quot;&gt;&lt;span style=&quot;top:-1.8478869999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.300005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.302113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4231360000000004em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.276864em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.321664em;vertical-align:-0.276864em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.19444em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4231360000000004em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.276864em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;굳이 오차제곱을 사용하는 이유가 있을까? 오차제곱이 아니라 절댓값(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;l&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L_1-loss&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.01968em&quot;&gt;l&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)을 사용할수는 없나?&lt;/p&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;실제로는 절댓값을 사용하는 것과 오차제곱을 사용하는 경우에 차이가 있다. 절댓값을 사용하는 방식을 &lt;strong&gt;&lt;code&gt;MAE(Mean Absolute Error)&lt;/code&gt;&lt;/strong&gt;라고 한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;MAE&lt;/code&gt;를 사용하는 경우&lt;ul&gt;&lt;li&gt;데이터에는 에러, outlier가 있기 마련이다. 그런데 에러가 있는 데이터가 너무 많다면, 그 데이터에 맞추려다가 모델 전반의 학습이 망가지게 된다.&lt;/li&gt;&lt;li&gt;MAE는 이 경우에 outlier에 대해 큰 페널티를 주지 않으므로, &lt;strong&gt;튀는 데이터값을 무시하고 강건한 모델링&lt;/strong&gt;을 할 수 있다.&lt;/li&gt;&lt;li&gt;다만, &lt;strong&gt;outlier가 너무 많다면 학습이 망가지게 된다.&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;MSE&lt;/code&gt;를 사용하는 경우&lt;ul&gt;&lt;li&gt;MAE에 비하여 MSE는 &lt;strong&gt;튀어나온 값(평균과 차이가 있는 값)에 더 큰 페널티&lt;/strong&gt;를 주므로, outlier들을 수용하는 모델링을 하기에 좋다.&lt;/li&gt;&lt;li&gt;절댓값보다 &lt;strong&gt;컴퓨터 연산이 더 쉽다.&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;정확한 설명은 다음을 참고하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://www.kaggle.com/c/home-data-for-ml-course/discussion/143364&quot;&gt;https://www.kaggle.com/c/home-data-for-ml-course/discussion/143364&lt;/a&gt;&lt;/p&gt;&lt;p&gt;이처럼 &lt;strong&gt;&lt;div&gt;상황에 따라 Loss Function을 다르게 사용&lt;/div&gt;&lt;/strong&gt;한다.&lt;/p&gt;&lt;h1 id=&quot;-7&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-7&quot; aria-label=&quot; 7 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이번엔, &lt;code&gt;Classification Task&lt;/code&gt;를 살펴보자.&lt;/p&gt;&lt;p&gt;분류문제에서는 주로 &lt;strong&gt;&lt;code&gt;Cross Entropy&lt;/code&gt;&lt;/strong&gt;가 Loss Function으로  사용된다. 굳이 크로스엔트로피를 사용하는 이유가 있을까?&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mfrac&gt;&lt;msubsup&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/msubsup&gt;&lt;msubsup&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;/msubsup&gt;&lt;msubsup&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;msubsup&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;CE = -\frac{1}{N}\sum^N_{i=1}\sum^D_{d=1}y_i^{(d)}\log \hat{y}_i^{(d)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.3898em;vertical-align:-0.345em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.845108em&quot;&gt;&lt;span style=&quot;top:-2.6550000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.394em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.345em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.981231em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2029em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29971000000000003em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.981231em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2029em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29971000000000003em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4231360000000004em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.27686399999999994em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.19444em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4231360000000004em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.27686399999999994em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/p&gt;&lt;p&gt;분류문제에서 정답레이블은 일반적으로 &lt;strong&gt;원-핫벡터&lt;/strong&gt;로 표기된다. 즉 크로스 엔트로피 식에서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;D&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개의  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y_i^{(d)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.321664em;vertical-align:-0.27686399999999994em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4231360000000004em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.27686399999999994em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;중 유효한 값은 하나밖에 없다. 그러므로, &lt;strong&gt;&lt;div&gt;내 모델의 출력값 중에서 정답과 일치하는 클래스의 값만 높여주겠다&lt;/div&gt;&lt;/strong&gt;는 의미이다. 얼마나 높일지가 중요한 게 아니라, 해당 차원의 값만 키워주는것이 목적이 된다.&lt;/p&gt;&lt;h1 id=&quot;-8&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-8&quot; aria-label=&quot; 8 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;마지막으로, &lt;code&gt;Probability Task&lt;/code&gt;를 살펴보자.&lt;/p&gt;&lt;p&gt;확률문제의 Loss Function으로는 &lt;strong&gt;&lt;code&gt;MLE(Maximum Likelihood Estimation)&lt;/code&gt;&lt;/strong&gt;가 많이 쓰인다.&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mfrac&gt;&lt;msubsup&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/msubsup&gt;&lt;msubsup&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;D&lt;/mi&gt;&lt;/msubsup&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msubsup&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo separator=&quot;true&quot;&gt;;&lt;/mo&gt;&lt;msubsup&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msubsup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mtext&gt;  &lt;/mtext&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;MLE = \frac{1}{N}\sum^N_{i=1}\sum^D_{d=1}\log \mathcal{N}(y_i^{(d)};\hat{y}_i^{(d)},1) \ \ (=MSE)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.3898em;vertical-align:-0.345em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.845108em&quot;&gt;&lt;span style=&quot;top:-2.6550000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.394em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.345em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.981231em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2029em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29971000000000003em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.981231em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2029em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29971000000000003em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot; style=&quot;margin-right:0.14736em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4231360000000004em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.27686399999999994em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.19444em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0448em&quot;&gt;&lt;span style=&quot;top:-2.4231360000000004em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.27686399999999994em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/p&gt;&lt;p&gt;예를 들어 회귀 문제를 풀었을 때, output을 단순히 출력값이 아닌 확률로 내보내는 모델을 만들고 싶다면 MLE를 사용한다. &lt;/p&gt;&lt;p&gt;가장 높은 가능성을 가진 예측 모델을 찾아야하기 때문이다.&lt;/p&gt;&lt;p&gt;만약 단순히 최고가능성만 따지는게 아니라, 해당 판단이 얼마나 확률적으로 높고, 얼마나 신뢰성이 있는지 등의 정보가 필요하다면 다른 손실함수를 사용하기도 한다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[Matplotlib와 Seaborn을 이용한 데이터 시각화]]></title><description><![CDATA[데이터 시각화 by 최성철 교수님, BoostCamp AI Tech 2주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/09_data_visualization/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/09_data_visualization/</guid><pubDate>Fri, 29 Jan 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 수강한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;데이터-시각화&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%8D%B0%EC%9D%B4%ED%84%B0-%EC%8B%9C%EA%B0%81%ED%99%94&quot; aria-label=&quot;데이터 시각화 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;데이터 시각화&lt;/h1&gt;&lt;h2 id=&quot;matplotlib&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#matplotlib&quot; aria-label=&quot;matplotlib permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;matplotlib&lt;/h2&gt;&lt;p&gt;파이썬의 대표적인 시각화도구로, 다양한 그래프를 지원하며 &lt;code&gt;pandas&lt;/code&gt;를 연동하므로 쉽게 사용할 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;pyplot&lt;/code&gt; 객체(일종의 그림판)를 사용하여 데이터를 표시한다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;pyplot&lt;/code&gt; 객체에 &lt;code&gt;figure&lt;/code&gt;들을 쌓은 다음, &lt;code&gt;show&lt;/code&gt;를 통해 show하고 flush(메모리에서 뱉어냄)한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;matplotlib&lt;/code&gt;&lt;/strong&gt;의 큰 단점은, argument를 kwargs로 받는다는 것이다.&lt;/p&gt;&lt;p&gt;고정된 argument가 없거나 문서화가 잘 안되어있어서, 매번 &lt;code&gt;shift&lt;/code&gt;+&lt;code&gt;tap&lt;/code&gt;을 눌러 argument들을 확인 후 사용해야한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;그냥 매번 구글링하는것을 추천한다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;figure--axes&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#figure--axes&quot; aria-label=&quot;figure  axes permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Figure &amp;amp; Axes&lt;/h3&gt;&lt;p&gt;Matplotlib는 &lt;strong&gt;&lt;code&gt;Axes&lt;/code&gt;&lt;/strong&gt;를 담은 &lt;strong&gt;&lt;code&gt;Figure&lt;/code&gt;&lt;/strong&gt;를 &lt;code&gt;pyplot&lt;/code&gt; 객체위에 올려서 표현한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;Figure&lt;/code&gt;를 구성할 때에는 x값의 순서를 고려해주어야한다. x값에 다라 따라 순차적으로 y값이 출력되기 때문이다. 만약 x값이 뒤죽박죽으로 섞여있다면, 선이 마구잡이로 그려진다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;code&gt;Axes&lt;/code&gt;는 subplot을 그리는 개념이라고 생각하면 된다.&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;bash&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-bash&quot;&gt;BASH&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-bash&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;fig &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; plt.figure&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# pyplot 위에 figure 객체를 쌓고 이를 반환함.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;fig.set_size_inches&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;5,2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 크기 지정&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;ax_1 &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; fig.add_subplot&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1,2&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;,1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 파라미터는 앞에서부터 row, column, figure의 순서&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;ax_2 &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; fig.add_subplot&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1,2&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;,2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;ax_1.plot&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;X_1, Y_1, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;b&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# c -&amp;gt; color&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;ax_2.plot&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;X_2, Y_2, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;g&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;h3 id=&quot;기초-사용법&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B8%B0%EC%B4%88-%EC%82%AC%EC%9A%A9%EB%B2%95&quot; aria-label=&quot;기초 사용법 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;기초 사용법&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;&lt;code&gt;plt.plot&lt;/code&gt; 메서드 파라미터&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;color&lt;/code&gt; , &lt;code&gt;c&lt;/code&gt;: 색상 지정&lt;ul&gt;&lt;li&gt;hex코드나 영어로 색상을 지정해 줄 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;linestyle&lt;/code&gt;, &lt;code&gt;ls&lt;/code&gt; : 선 형태 지정&lt;ul&gt;&lt;li&gt;&lt;code&gt;dashed&lt;/code&gt;, &lt;code&gt;dashdot&lt;/code&gt; 등의 옵션이 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;code&gt;plt.title&lt;/code&gt; : 플롯에 제목 추가&lt;/p&gt;&lt;ul&gt;&lt;li&gt;Latex 형식의 표현도 가능하다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;code&gt;plt.legend&lt;/code&gt; : 범례 표시&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;shadow&lt;/code&gt;, &lt;code&gt;fancybox&lt;/code&gt;, &lt;code&gt;loc&lt;/code&gt; 등 지정&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;code&gt;plt.grid&lt;/code&gt; : 보조선을 그음&lt;/p&gt;&lt;ul&gt;&lt;li&gt;투명도, 선 모양 등을 설정할 수 있다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;plt.style.use(&amp;quot;ggplot&amp;quot;)&lt;/code&gt; : &lt;code&gt;R&lt;/code&gt;에서 많이 쓰이는 형태의 grid 표시&lt;ul&gt;&lt;li&gt;이외에도 style은 다양하니 찾아보자.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;code&gt;plt.xlim&lt;/code&gt; : x축 범위 한계 지정&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;code&gt;plt.ylim&lt;/code&gt; : y축 범위 한계 지정&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;code&gt;plt.text&lt;/code&gt; : plot에 텍스트 추가&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;plt.annotate&lt;/code&gt; : 텍스트 추가하고, 화살표를 표시 할 수 있다&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;code&gt;plt.savefig(&amp;#x27;file_name.png&amp;#x27;)&lt;/code&gt; : 특정 파일명으로 figure 저장&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;show&lt;/code&gt;로 flush 하기 전에 저장할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;matplotlib-graph&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#matplotlib-graph&quot; aria-label=&quot;matplotlib graph permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;matplotlib graph&lt;/h2&gt;&lt;h3 id=&quot;scatter&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#scatter&quot; aria-label=&quot;scatter permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;scatter&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;산점도&lt;/strong&gt;를 표시한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;X, Y 데이터가 파라미터로 들어가지만, Figure를 만들때처럼 X 값의 순서가 중요하지는 않다.&lt;/li&gt;&lt;li&gt;마커의 형태를 선택할 수 있으니, 구글링해서 찾아보자.&lt;/li&gt;&lt;li&gt;&lt;code&gt;s&lt;/code&gt; 옵션으로 데이터 크기를 지정하고, 데이터 크기를 비교할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;bash&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-bash&quot;&gt;BASH&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-bash&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;data_1 &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; np.random.rand&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;512&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;data_2 &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; np.random.rand&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;512&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;plt.scatter&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;data_1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;:,0&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, data_1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;:,1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;b&amp;quot;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;marker&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;x&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;plt.scatter&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;data_2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;:,0&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, data_2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;:,1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;c&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;r&amp;quot;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;marker&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;o&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 한 subplot 위에 여러 scatter를 쌓을수도 있다.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;plt.show&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:758px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:64.0625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;scatter&quot; title=&quot;scatter&quot; src=&quot;/static/8e61b25bbd88edb79de829e03efa9e7a/c8e86/scatter.png&quot; srcSet=&quot;/static/8e61b25bbd88edb79de829e03efa9e7a/6f3f2/scatter.png 256w,/static/8e61b25bbd88edb79de829e03efa9e7a/01e7c/scatter.png 512w,/static/8e61b25bbd88edb79de829e03efa9e7a/c8e86/scatter.png 758w&quot; sizes=&quot;(max-width: 758px) 100vw, 758px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;bar-chart&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#bar-chart&quot; aria-label=&quot;bar chart permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;bar chart&lt;/h3&gt;&lt;p&gt;&lt;code&gt;bar&lt;/code&gt; 함수를 사용하여, 막대 차트를 그린다.&lt;/p&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;bash&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-bash&quot;&gt;BASH&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-bash&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;data &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;., &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;25&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;., &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;50&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;., &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;20&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;,&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;        &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;4&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;., &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;23&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;., &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;51&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;., &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;17&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;,&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;        &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;6&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;., &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;22&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;., &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;52&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;., &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;19&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;X &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; np.arange&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;0,8&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;,2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;plt.bar&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;X + &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;0.00&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, data&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, color &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;b&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, width &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;0.50&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 폭이 0.5&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;plt.bar&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;X + &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;0.50&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, data&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, color &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;g&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, width &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;0.50&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# x축 시작위치로, 이전 bar만큼의 폭은 건너뜀&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;plt.bar&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;X + &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1.0&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, data&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, color &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;r&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, width &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;0.50&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;plt.xticks&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;X+0.50, &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;A&amp;quot;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;B&amp;quot;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;C&amp;quot;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;D&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;))&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 시퀀스 타입으로 넣어야 들어간다.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;plt.show&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 마찬가지로, 막대차트의 y축 방향으로 여러 차트를 쌓아 보여줄수도 있다.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 여러 variation은 구글링해보기.&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:766px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:64.453125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;bar&quot; title=&quot;bar&quot; src=&quot;/static/4cf2199093348fd7f14d695e1c2d039a/f7616/bar.png&quot; srcSet=&quot;/static/4cf2199093348fd7f14d695e1c2d039a/6f3f2/bar.png 256w,/static/4cf2199093348fd7f14d695e1c2d039a/01e7c/bar.png 512w,/static/4cf2199093348fd7f14d695e1c2d039a/f7616/bar.png 766w&quot; sizes=&quot;(max-width: 766px) 100vw, 766px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;histogram&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#histogram&quot; aria-label=&quot;histogram permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;histogram&lt;/h3&gt;&lt;p&gt;도수분포표를 그림으로 나타내주는 히스토그램을 그린다.&lt;/p&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;bash&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-bash&quot;&gt;BASH&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-bash&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;X &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; np.random.normal&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;0,100&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;,1000&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;plt.hist&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;X, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;bins&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;20&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# bins는 꺾인 구간.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;plt.show&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:766px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:64.453125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAANCAIAAAAmMtkJAAAACXBIWXMAABYlAAAWJQFJUiTwAAAB5UlEQVQoz22SXUvbUBiAe7E7ZdILP/DjYmwMV73y3v/hlT9gF25WW6ejFgUvvBUEQfQHWKZrm7boFBQG29iFRrQurjGirW1o0iYnn03SxDeNOw7Ny+HlnMN5zvPmzQk4jtNoNERRVBRF/i9EhFheEJEk+4UkSaqqBgRBYFnWsiznX9i2DblYV6cSpLt0/ANkAU3TgDdN8xFu5dnt07ex3K8rHuZW0/aHoYZarYbhZku7dkR3Tqc6Jr+OLH37U0Z4/ymMEOJ5HpftHQonyPaJnTex3Iv3X1YPC75yF9Z1HWBs9g7FkmftH3YGZjO9M8RgfJeq+Mhd2DAMDHutqkr62PrPYDjZ/ynTN0NA/Se3dX8Y/hOGzZZ26/dN28SD1htnJdGFm89gwHDDvLtXDi6Dk64WMDB3TafCiRPTsm3bF+Z5wzStpru1n6/0RNLYCXBPNN0dSW9+Z7yO4BsevpnjOMNwzeStMLSwFwynuqJEZyTtje4o8fJjcnhxj67KuC8Q7guDBbxNJCkZsji6vP/6MxGKZ0Pz2Xfz2cFYBnIonhuO517NEeMbP44uyiVeUvUGUjQVzABrmvq3QOeZuzx9Q9FMmauT5xRbE69LFU6ULwoMU6wU2erx+eU1K5zmKUlWqQItCPV7lauT7nu3XgQAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;histogram&quot; title=&quot;histogram&quot; src=&quot;/static/8a6b9e2f11a4165bb1aaeaac7aff5972/f7616/histogram.png&quot; srcSet=&quot;/static/8a6b9e2f11a4165bb1aaeaac7aff5972/6f3f2/histogram.png 256w,/static/8a6b9e2f11a4165bb1aaeaac7aff5972/01e7c/histogram.png 512w,/static/8a6b9e2f11a4165bb1aaeaac7aff5972/f7616/histogram.png 766w&quot; sizes=&quot;(max-width: 766px) 100vw, 766px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;값들의 분포상태를 알 수 있는 박스플롯을 그릴수도 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;박스는, 데이터가 몰려있는 구간을 나타낸다.&lt;/li&gt;&lt;li&gt;선은, 어느정도의 데이터가 분포하는 구간을 나타낸다.(박스의 길이에 일정량을 곱하여 길이를 정한다)&lt;/li&gt;&lt;li&gt;점은, outlier를 의미한다.&lt;/li&gt;&lt;/ul&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;bash&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-bash&quot;&gt;BASH&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-bash&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;data &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; np.random.randn&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;100,5&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;plt.boxplot&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;data&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;plt.show&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:762px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:66.796875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;boxplot&quot; title=&quot;boxplot&quot; src=&quot;/static/37ed40fe0c358400b5318e0f665c65f7/a016c/boxplot.png&quot; srcSet=&quot;/static/37ed40fe0c358400b5318e0f665c65f7/6f3f2/boxplot.png 256w,/static/37ed40fe0c358400b5318e0f665c65f7/01e7c/boxplot.png 512w,/static/37ed40fe0c358400b5318e0f665c65f7/a016c/boxplot.png 762w&quot; sizes=&quot;(max-width: 762px) 100vw, 762px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;seaborn&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#seaborn&quot; aria-label=&quot;seaborn permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;seaborn&lt;/h2&gt;&lt;p&gt;통계 데이터 시각화 도구로, &lt;code&gt;matplotlib&lt;/code&gt;를 좀 더 쉽고 다루기 위해 지원해주는 wrapper library이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;matplotlib와 같은 기본적인 plot을 지원한다.&lt;/li&gt;&lt;li&gt;손쉬운 설정으로 데이터를 (예쁘게) 산출한다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;basic plot&lt;/code&gt;과 &lt;code&gt;multi plot&lt;/code&gt;이 있다.&lt;ul&gt;&lt;li&gt;basic plot&lt;ul&gt;&lt;li&gt;&lt;code&gt;lineplot&lt;/code&gt;, &lt;code&gt;scatterplot&lt;/code&gt;, &lt;code&gt;countplot&lt;/code&gt; 등&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;multi plot&lt;ul&gt;&lt;li&gt;주로 많이 사용한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;a href=&quot;https://seaborn.pydata.org/tutorial.html&quot;&gt;튜토리얼&lt;/a&gt;이 굉장히 잘되어있는 편이라 읽어보는 것을 추천한다.&lt;/p&gt;&lt;h3 id=&quot;basic-plot&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#basic-plot&quot; aria-label=&quot;basic plot permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;basic plot&lt;/h3&gt;&lt;p&gt;&lt;code&gt;matplotlib&lt;/code&gt;에 비해 사용법이 간단하다.&lt;/p&gt;&lt;p&gt;x와 y 파라미터에 직접 시퀀스형 자료를 넣어도 되지만, 최근에는 &lt;code&gt;pandas&lt;/code&gt; 자료형도 지원한다.&lt;/p&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;bash&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-bash&quot;&gt;BASH&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-bash&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token function&quot;&gt;import&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; numpy as np&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token function&quot;&gt;import&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; pandas as pd&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token function&quot;&gt;import&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; matplotlib.pyplot as plt&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token function&quot;&gt;import&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; seaborn as sns&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;sns.set&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;style&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;darkgrid&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;tips &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; sns.load_dataset&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;tips&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;fmri &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; sns.load_dataset&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;fmri&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# line plot&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# x : x축 이름, y : y축 이름, data : pandas 데이터&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# x축이 정렬되어있지 않아도 정렬해서 보여준다.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;sns.lineplot&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;timepoint&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;signal&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;data&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;fmri&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# hue 라는 파라미터를 넣어 catergory별로 나누어서 데이터를 시각화할수도 있다.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 아래는, &amp;#x27;event&amp;#x27; 항목에 들어가있는 값들을 카테고리별로 분류해서 시각화하겠다는 것.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;sns.lineplot&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;timepoint&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;signal&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;hue&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;event, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;data&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;fmri&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# count plot - 카테고리형 데이터를 위해 사용&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;sns.coutplot&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;smoker&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;hue&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;time&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;data&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;tips&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# bar plot - estimator로 원하는 통계치를 시각화 할수있음.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;sns.barplot&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;day&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;total_bill&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;data&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;tips, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;estimator&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;np.std&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;h3 id=&quot;multi-plot&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#multi-plot&quot; aria-label=&quot;multi plot permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;multi plot&lt;/h3&gt;&lt;p&gt;이외에도 여러 plot이 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;Violinplot&lt;/code&gt; - &lt;code&gt;boxplot&lt;/code&gt;에 distribution을 함께 표현한다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;Stripplot&lt;/code&gt; - scatter와 category 정보를 함께 표현한다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;Swarmplot&lt;/code&gt; - 분포와 scatter를 표현한다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;Pointplot&lt;/code&gt; - 카테고리별로 numeri의 평균, 신뢰구간을 표시한다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;regplot&lt;/code&gt; - scatter과 (회귀)선형함수를 함께 표시한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;또, &lt;code&gt;FacetGrid&lt;/code&gt; 로 카테고리데이터를 결합분포로 볼 수 있다.&lt;/p&gt;</content:encoded></item><item><title><![CDATA[인공지능 통계학 기초]]></title><description><![CDATA[통계학 맛보기 by 임성빈, BoostCamp AI Tech 2주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/10_statistics/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/10_statistics/</guid><pubDate>Fri, 29 Jan 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 수강한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;통계학-맛보기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%86%B5%EA%B3%84%ED%95%99-%EB%A7%9B%EB%B3%B4%EA%B8%B0&quot; aria-label=&quot;통계학 맛보기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;통계학 맛보기&lt;/h1&gt;&lt;h2 id=&quot;통계적-모델링&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%86%B5%EA%B3%84%EC%A0%81-%EB%AA%A8%EB%8D%B8%EB%A7%81&quot; aria-label=&quot;통계적 모델링 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;통계적 모델링&lt;/h2&gt;&lt;p&gt;적절한 가정 위에서 확률 분포를 추정하는 것&lt;/p&gt;&lt;p&gt;확률 분포 종류도 아주 다양&lt;/p&gt;&lt;p&gt;그러나 유한한 개수의 데이터만 관찰해서 모집단의 분포를 정확하게 알아낸다는 것은 불가능하다. 근사적으로 확률분포를 추정할 수 밖에 없다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;불확실성을 고려한 상태에서 예측의 위험성을 최소화하는 방향으로 추정한다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;모수&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%AA%A8%EC%88%98&quot; aria-label=&quot;모수 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;모수&lt;/h2&gt;&lt;h3 id=&quot;모수란&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%AA%A8%EC%88%98%EB%9E%80&quot; aria-label=&quot;모수란 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;모수란?&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;모수(Parameter)&lt;/code&gt;&lt;/strong&gt;란, &lt;strong&gt;모집단의 특성을 나타내주는 통계치&lt;/strong&gt;이다. 우리가 통계학을 사용할 때에는 모집단을 전수조사 할 수 없기 때문에 표본조사를 하는데, 이 때 표본조사를 통해 모수를 추정함으로써 모집단의 특성을 추측할 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;모집단이 어떤 확률분포를 따르느냐에 따라 각기 다른 통계치들을 모수로 사용한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;데이터가 특정 확률 분포를 따른다고 선험적으로(a priori) 가정한 후 그 분포를 결정하는 모수(parameter)를 추정하는 방법을 &lt;strong&gt;&lt;code&gt;모수적(parametric)&lt;/code&gt;&lt;/strong&gt; 방법론이라고 한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;예를 들어 정규분포를 가지고 확률적 모델링을 하면, 가장 중요한 평균과 분산을 묶어서 모수라고 부른다.&lt;ul&gt;&lt;li&gt;이 때, 평균과 분산을 추정하기 위해 학습하는 방법을 &lt;code&gt;모수적 방법론&lt;/code&gt;이라고 한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;반면, 특정 확률 분포를 가정하지 않고 데이터에 따라 모델의 구조 및 모수의 개수가 유연하게 바뀌면 &lt;strong&gt;&lt;code&gt;비모수(nonparametric)&lt;/code&gt;&lt;/strong&gt; 방법론이라고 부른다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;명심할점은, &lt;code&gt;비모수 방법론&lt;/code&gt;이라고 해서 모수가 없다거나 쓰지 않는 것이 아니라는 것이다.&lt;ul&gt;&lt;li&gt;모수가 무한히 많거나, 모수의 수가 데이터에 따라 변할 수 있는 경우이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;모수적 방법론&lt;/code&gt;과 &lt;code&gt;비모수 방법론&lt;/code&gt;의 차이는 &lt;div&gt;확률분포에 대한 가정 부여 여부&lt;/div&gt;이다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;확률분포-가정하기--예제&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%99%95%EB%A5%A0%EB%B6%84%ED%8F%AC-%EA%B0%80%EC%A0%95%ED%95%98%EA%B8%B0--%EC%98%88%EC%A0%9C&quot; aria-label=&quot;확률분포 가정하기  예제 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;확률분포 가정하기 : 예제&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;우선, 히스토그램을 통해 데이터의 모양을 관찰한다.&lt;ul&gt;&lt;li&gt;데이터가 2개(0 또는 1)의 값을 가지는 경우 → &lt;code&gt;베르누이분포&lt;/code&gt;&lt;/li&gt;&lt;li&gt;데이터가 n개의 이산적인 값을 가지는 경우 → &lt;code&gt;카테고리분포&lt;/code&gt;&lt;/li&gt;&lt;li&gt;데이터가 [0,1] 사이에서 값을 가지는 경우 → &lt;code&gt;베타분포&lt;/code&gt;&lt;/li&gt;&lt;li&gt;데이터가 0 이상의 값을 가지는 경우 → &lt;code&gt;감마분포&lt;/code&gt;, &lt;code&gt;로그정규분포&lt;/code&gt; 등&lt;/li&gt;&lt;li&gt;데이터가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;R&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathbb{R}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68889em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 전체에서 값을 가지는 경우 → &lt;code&gt;정규분포&lt;/code&gt;, &lt;code&gt;라플라스분포&lt;/code&gt; 등&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;단, 위의 기준에 따라 기계적으로 확률분포를 가정해서는 안된다. &lt;strong&gt;&lt;div&gt;데이터를 생성하는 원리를 먼저 고려하는것이 원칙&lt;/div&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이는 통계학과 기계학습 모두에서 강조하고 있는 점이다.&lt;/li&gt;&lt;li&gt;각 분포마다 검정하는 방법들이 있으므로, 모수를 추정한 후에는 반드시 &lt;strong&gt;통계적 검정&lt;/strong&gt;을 해야 한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;확률 분포들의 자세한 설명에 대해서는 다음 글을 참고해보자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://yeomko.tistory.com/33&quot;&gt;갈아먹는 통계 기초 [1] 확률 분포 정리&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;모수-추정하기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%AA%A8%EC%88%98-%EC%B6%94%EC%A0%95%ED%95%98%EA%B8%B0&quot; aria-label=&quot;모수 추정하기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;모수 추정하기&lt;/h3&gt;&lt;p&gt;데이터로 확률분포를 가정했다면, 모수를 추정해 볼 수 있다.&lt;/p&gt;&lt;p&gt;예를 들어 정규분포의 모수는 평균 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mu&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 과 분산 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sigma^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8141079999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 있다. 이를 추정하는 통계량(statistic)은 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mtext&gt;표본평균 &lt;/mtext&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo&gt;ˉ&lt;/mo&gt;&lt;/mover&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mfrac&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/munderover&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo&gt;ˉ&lt;/mo&gt;&lt;/mover&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;표본평균 \ \bar{X} = \frac{1}{N}\displaystyle\sum^N_{i=1}X_i\\
\mathbb{E}[\bar{X}] = \mu&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8201099999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;표&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;본&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;평&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;균&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8201099999999999em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.25233em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.16666em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;ˉ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.106005em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.8283360000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.07011em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8201099999999999em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.25233em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.16666em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;ˉ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mtext&gt;표본분산 &lt;/mtext&gt;&lt;msup&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/munderover&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo&gt;ˉ&lt;/mo&gt;&lt;/mover&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;표본분산 \ S^2 = \frac{1}{N-1}\displaystyle\sum^N_{i=1}(X_i-\bar{X})^2\\
\mathbb{E}[S^2] = \sigma^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8641079999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;표&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;본&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;분&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;산&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.106005em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7693300000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.8283360000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.07847em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1141079999999999em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8201099999999999em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.25233em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.16666em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;ˉ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1141079999999999em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8641079999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;주의할 점은, 표본분산에서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;/&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;1/(N-1)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;/&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 사용하는것. 이는 불편(unbiased) 추정량을 구하기 위함이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;기댓값을 취했을 때 원래 모집단을 대표하는 통계값에 일치하도록 사용한다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;왜 표본분산은 모분산을 구할때처럼 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;으로 나누지 않는가가 이해되지 않는다면, 수식없이 풀어놓은 다음 동영상을 참고한다.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://www.youtube.com/watch?v=frz-BE3a6H0&amp;amp;feature=emb_logo&amp;amp;ab_channel=ASDF%EC%98%A4%ED%84%B0%EC%9D%98%ED%86%B5%EA%B3%84&quot;&gt;수식없음!!!왜 표본분산은 n아 아니라 n-1로 나눌까?&lt;/a&gt;&lt;/p&gt;&lt;p&gt;또, 왜 정확히 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N-1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.76666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 나누는가에 대해서는 아래의 포스팅을 참고하도록 하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://m.blog.naver.com/sw4r/221021838997&quot;&gt;왜 표본(샘플)의 분산에서는 n이 아닌 n-1로 나눌까?&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이런 &lt;strong&gt;&lt;code&gt;통계량&lt;/code&gt;의 확률분포(즉, 표본평균과 표본 분산의 확률분포)&lt;/strong&gt;를 &lt;strong&gt;&lt;code&gt;표집분포(sampling distribution)&lt;/code&gt;&lt;/strong&gt;라고 부르며, 특히 &lt;strong&gt;&lt;code&gt;표본평균의 표집분포&lt;/code&gt;&lt;/strong&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 커질수록 정규분포 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;/&lt;/mi&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{N}(\mu,\sigma^2/N)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.064108em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.3525em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;/&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 따른다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이를 &lt;strong&gt;&lt;code&gt;중심극한정리(Central Limit Theorem)&lt;/code&gt;&lt;/strong&gt;이라 부르며, 모집단의 분포가 정규분포를 따르지 않아도 성립한다.&lt;ul&gt;&lt;li&gt;즉, 모집단의 분포가 정규분포가 아니더라도 &lt;strong&gt;&lt;code&gt;표본평균의 표집분포(sampling distribution)&lt;/code&gt;&lt;/strong&gt;는 정규분포를 따를 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;div&gt;&lt;code&gt;표본분포(sample distribution)&lt;/code&gt;와 &lt;code&gt;표집분포(sampling distribution)&lt;/code&gt;는 다르다&lt;/div&gt;&lt;/strong&gt;는 점에 주의하자.&lt;ul&gt;&lt;li&gt;원래 모집단의 분포가 정규분포를 따르지 않는다면, &lt;strong&gt;&lt;code&gt;표본분포&lt;/code&gt;는 N이 아무리 커져도 (당연히) 정규분포를 따를 수 없다.&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;그러나 &lt;strong&gt;&lt;code&gt;표본평균의 표집분포&lt;/code&gt;&lt;/strong&gt;는 N이 커짐에 따라 정규분포를 따르게 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:74.21875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;sampling_distribution&quot; title=&quot;sampling_distribution&quot; src=&quot;/static/d9f424fb5ca5636b3882c49d43ab9c1e/2bef9/sampling_distribution.png&quot; srcSet=&quot;/static/d9f424fb5ca5636b3882c49d43ab9c1e/6f3f2/sampling_distribution.png 256w,/static/d9f424fb5ca5636b3882c49d43ab9c1e/01e7c/sampling_distribution.png 512w,/static/d9f424fb5ca5636b3882c49d43ab9c1e/2bef9/sampling_distribution.png 1024w,/static/d9f424fb5ca5636b3882c49d43ab9c1e/d74fe/sampling_distribution.png 1164w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;위의 사진은 베르누이확률분포(이항확률분포)를 따르는 데이터에서, 표본평균의 표집분포를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 늘어남에 따라 기록한것이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;가장 왼쪽은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 적을때인데, 표본평균의 분포가 양극단으로 찍히는 경우도 있다.&lt;ul&gt;&lt;li&gt;예를 들어 [1,2,3,4,5,6,7,8,9,10] 중 [1,2,3], [8,9,10]이라는 표본들을 골라서 표본 평균을 찍어보면, 2와 9로 평균의 분포가 양극단으로 찍힐 것이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;그러나, &lt;strong&gt;&lt;div&gt;N이 늘어날수록 점점 표본평균의 분포는 중간으로 모이다가, 결국은 정규분포를 따라간다.&lt;/div&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;위의 예에서 [1,2,3],[8,9,10]이 아닌 &lt;strong&gt;새로운 표본들을 점점 추가하면 추가할수록, 이 표본들의 평균(표본평균)을 다시 평균내면 모집단의 평균인 5에 근접&lt;/strong&gt;하게 될것이다.&lt;ul&gt;&lt;li&gt;바꾸어 말하면, 표본평균의 분산(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mfrac&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{\sigma^2}{N})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.36292em;vertical-align:-0.345em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.01792em&quot;&gt;&lt;span style=&quot;top:-2.6550000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.394em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913142857142857em&quot;&gt;&lt;span style=&quot;top:-2.931em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.345em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 늘어날수록 0에 가까워진다.&lt;/li&gt;&lt;li&gt;이는 곧 표본평균들의 (표집)분포가 정규분포를 따른다는 것을 의미한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;즉, 원래 &lt;strong&gt;모집단의 확률분포는 정규분포가 아닌 이항분포&lt;/strong&gt;임에도 불구하고, &lt;strong&gt;표본평균의 확률분포(표집분포)는 정규분포를 따른다&lt;/strong&gt;는 것을 확인할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;모수와 중심극한정리의 개념이 잘 이해되지 않는다면 다음 글을 참고한다.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://chukycheese.github.io/data%20science/parameter-clt/&quot;&gt;모수, 큰 수의 법칙, 그리고 중심극한정리&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;http://blog.naver.com/PostView.nhn?blogId=khj90733&amp;amp;logNo=221904329150&amp;amp;parentCategoryNo=61&amp;amp;categoryNo=&amp;amp;viewDate=&amp;amp;isShowPopularPosts=true&amp;amp;from=search&quot;&gt;모집단분포, 표본분포, 표집분포 / 중심극한정리 / 분산, 표준편차의 편파추정치와 불편파 추정치 / 자유도&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-7&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-7&quot; aria-label=&quot; 7 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h1 id=&quot;-8&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-8&quot; aria-label=&quot; 8 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;최대가능도추정법maximum-likelihood-estimationmle&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%B5%9C%EB%8C%80%EA%B0%80%EB%8A%A5%EB%8F%84%EC%B6%94%EC%A0%95%EB%B2%95maximum-likelihood-estimationmle&quot; aria-label=&quot;최대가능도추정법maximum likelihood estimationmle permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;최대가능도추정법(Maximum Likelihood Estimation,MLE)&lt;/h2&gt;&lt;h3 id=&quot;최대가능도-추정법이란&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%B5%9C%EB%8C%80%EA%B0%80%EB%8A%A5%EB%8F%84-%EC%B6%94%EC%A0%95%EB%B2%95%EC%9D%B4%EB%9E%80&quot; aria-label=&quot;최대가능도 추정법이란 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;최대가능도 추정법이란?&lt;/h3&gt;&lt;p&gt;표본평균이나 표본분산은 중요한 통계랑이지만, 확률분포마다 사용하는 모수가 다르므로 적절한 통계량이 달라지게 된다.&lt;/p&gt;&lt;p&gt;따라서, 해당 표본으로부터 무엇을 모수로 추정할 것인지 정하는 방법이 필요하다.&lt;/p&gt;&lt;p&gt;이론적으로 가장 가능성이 높은 모수를 추정하는 방법 중 하나는 &lt;strong&gt;&lt;code&gt;최대가능도추정법(maximum likelihood estimation, MLE)&lt;/code&gt;&lt;/strong&gt;이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mrow&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;arg max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;;&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;arg max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\hat{\theta}_{MLE} = \underset{\theta}{\argmax}\ L(\theta;x) = \underset{\theta}{\argmax} \ P(x|\theta)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1078799999999998em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9578799999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.26344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.16666em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.696548em;vertical-align:-0.946548em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.153452em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.946548em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.696548em;vertical-align:-0.946548em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.153452em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.946548em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;위의 &lt;strong&gt;&lt;code&gt;가능도함수&lt;/code&gt; &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;;&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L(\theta;x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/strong&gt;는 주어진 데이터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해서(다시 말해, 데이터가 주어져 있는 상황에서) &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 바뀜에 따라 값이 바뀌는 함수이다.&lt;ul&gt;&lt;li&gt;모수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 따르는 분포가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 관찰할 가능성을 뜻한다.&lt;/li&gt;&lt;li&gt;확률론 기초에서 배웠던 확률[밀도/질량]함수들과 형태가 같다. 그러나, 관점에 차이가 있다.&lt;ul&gt;&lt;li&gt;원래 확률함수들은 모수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 가 주어져있을 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대한 함수로 생각한다. 그러나 가능도 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 은 모수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 변수로 둔 함수로 생각한다.&lt;/li&gt;&lt;li&gt;따라서 확률로 해석하면 안된다. &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sum&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.00001em;vertical-align:-0.25001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;하거나, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo&gt;∫&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\int&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.11112em;vertical-align:-0.30612em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;margin-right:0.19445em;position:relative;top:-0.0005599999999999772em&quot;&gt;∫&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;하였을 때 1이 나오는 개념이 아니다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해서 대소비교가 가능한 수를 추정하는 방식이라고 생각하자.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;;&lt;/mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mtext&gt;    &lt;/mtext&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo&gt;&amp;gt;&lt;/mo&gt;&lt;mtext&gt;    &lt;/mtext&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;;&lt;/mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L(\theta;X) = \prod^{n}_{i=1}P(x_i|\theta) \ \ \ \ =&amp;gt; \ \ \ \ \log L(\theta;X) = \displaystyle\sum^n_{i=1}\log P(x_i\theta)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.929066em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∏&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.5782em;vertical-align:-0.0391em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;&amp;gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.929066em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;데이터 집합 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 &lt;strong&gt;독립적으로 추출되었을 경우, &lt;code&gt;로그가능도&lt;/code&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log L&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 최적화&lt;/strong&gt;한다.&lt;ul&gt;&lt;li&gt;가능도 함수에서, 만약 데이터 집합 X의 각 행벡터(즉, 각 데이터)가 독립적으로 추출되었을 경우, 확률[밀도/질량]함수들의 곱(product)으로 표현할 수 있다.&lt;/li&gt;&lt;li&gt;이 경우 곱셈을 덧셈으로 바꾸어주는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 함수의 성질을 이용해서, &lt;strong&gt;확률분포함수들의 곱셈을 덧셈으로 바꿀 수 있다.&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;그렇다면, 원래 &lt;code&gt;가능도&lt;/code&gt; &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 를 최적화하는 모수나 &lt;code&gt;로그가능도&lt;/code&gt; &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log L&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 을 최적화하는 모수나 똑같이 MLE가 되는데, 왜 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log L&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 사용해야 할까?&lt;/p&gt;&lt;ul&gt;&lt;li&gt;데이터의 숫자가 적으면 상관 없지만, 데이터 숫자가 수억단위가 된다면 컴퓨터 정확도로는 &lt;code&gt;로그가능도&lt;/code&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log L&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 계산하는 것이 불가능하다.&lt;ul&gt;&lt;li&gt;가령 0에서 1사이의 확률값을 수억번 곱해준다고 생각해보자. 컴퓨터의 자릿수 표현에 관한 현실적 한계때문에, 연산오차가 나게 된다.&lt;/li&gt;&lt;li&gt;만약 데이터가 독립이라 &lt;code&gt;로그가능도&lt;/code&gt;를 사용할 수 있다면, &lt;code&gt;가능도&lt;/code&gt;의 곱셈(ex - 0과 1 사이의 수억번의 곱셈)을 &lt;code&gt;로그가능도&lt;/code&gt;의 덧셈으로 바꿀 수 있다.&lt;ul&gt;&lt;li&gt;따라서 &lt;strong&gt;컴퓨터로 연산이 가능한 수&lt;/strong&gt;가 되기때문에, &lt;strong&gt;최적화가 가능&lt;/strong&gt;해진다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;또, &lt;code&gt;경사하강법&lt;/code&gt;으로 가능도를 최적화할 때 &lt;strong&gt;미분 연산&lt;/strong&gt;을 사용하게 되는데, ****&lt;code&gt;로그가능도&lt;/code&gt;를 사용하면 연산량을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O(n^2)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.064108em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O(n)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;으로 줄여준다.&lt;ul&gt;&lt;li&gt;연산의 복잡도가 선형적으로 변하므로, &lt;strong&gt;효율적인 연산&lt;/strong&gt;이 가능해지며, &lt;strong&gt;연산의 오차범위 내에서 계산이 가능&lt;/strong&gt;해진다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;따라서, &lt;strong&gt;&lt;div&gt;로그 가능도를 사용하는 문제는 최적화 문제와 깊은 연관&lt;/div&gt;&lt;/strong&gt;이 있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;대개 손실함수의 경우는 &lt;code&gt;경사하강법&lt;/code&gt;을 사용하므로, 목적식을 최대화하는것이 아니라 &lt;strong&gt;최소화&lt;/strong&gt;한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;그래서 보통 최소화하는 데에 사용하기 위해서 그냥 &lt;code&gt;로그가능도&lt;/code&gt;가 아니라 &lt;strong&gt;&lt;code&gt;음의 로그가능도(negative log-likehood)&lt;/code&gt;&lt;/strong&gt;를 최적화하게 된다.&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;최대가능도mle-추정법-예제--정규분포&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%B5%9C%EB%8C%80%EA%B0%80%EB%8A%A5%EB%8F%84mle-%EC%B6%94%EC%A0%95%EB%B2%95-%EC%98%88%EC%A0%9C--%EC%A0%95%EA%B7%9C%EB%B6%84%ED%8F%AC&quot; aria-label=&quot;최대가능도mle 추정법 예제  정규분포 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;최대가능도(MLE) 추정법 예제 : 정규분포&lt;/h3&gt;&lt;p&gt;&lt;code&gt;정규분포&lt;/code&gt;를 따르는 확률변수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로부터 독립적인 표본 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;{&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;}&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\{x_1,\dots,x_n\}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;}&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 얻었을 때, &lt;strong&gt;&lt;code&gt;최대가능도추정법&lt;/code&gt;&lt;/strong&gt;을 이용해서 모수를 추정해보자.&lt;/p&gt;&lt;p&gt;먼저, &lt;code&gt;정규분포&lt;/code&gt;이기 때문에 두개의 모수 &lt;code&gt;평균&lt;/code&gt;과 &lt;code&gt;분산&lt;/code&gt;을 사용한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mrow&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;arg max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;;&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;arg max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\hat{\theta}_{MLE} = \underset{\theta}{\argmax}\ L(\theta;x) = \underset{\mu,\sigma^2}{\argmax} \ P(X|\mu, \sigma^2)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1078799999999998em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9578799999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.26344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.16666em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.696548em;vertical-align:-0.946548em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.153452em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.946548em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.983076em;vertical-align:-1.1189680000000002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.11714em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;μ&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7463142857142857em&quot;&gt;&lt;span style=&quot;top:-2.786em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1189680000000002em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;가능도 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 최적화하는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 찾는 문제이다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;;&lt;/mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;msqrt&gt;&lt;mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;π&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/msqrt&gt;&lt;/mfrac&gt;&lt;msup&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mrow&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log L(\theta;X) = \displaystyle\sum^n_{i=1}\log P(x_i|\theta) = \sum^n_{i=1}\log\frac{1}{\sqrt{2\pi\sigma^2}}e^{-\frac{\vert x_i-\mu\vert^2}{2\sigma^2}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.929066em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.929066em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.154946em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord sqrt&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9550540000000001em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;padding-left:0.833em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;π&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.740108em&quot;&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.915054em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail&quot; style=&quot;min-width:0.853em;height:1.08em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;1.08em&quot; viewBox=&quot;0 0 400000 1080&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M95,702
c-2.7,0,-7.17,-2.7,-13.5,-8c-5.8,-5.3,-9.5,-10,-9.5,-14
c0,-2,0.3,-3.3,1,-4c1.3,-2.7,23.83,-20.7,67.5,-54
c44.2,-33.3,65.8,-50.3,66.5,-51c1.3,-1.3,3,-2,5,-2c4.7,0,8.7,3.3,12,10
s173,378,173,378c0.7,0,35.3,-71,104,-213c68.7,-142,137.5,-285,206.5,-429
c69,-144,104.5,-217.7,106.5,-221
l0 -0
c5.3,-9.3,12,-14,20,-14
H400000v40H845.2724
s-225.272,467,-225.272,467s-235,486,-235,486c-2.7,4.7,-9,7,-19,7
c-6,0,-10,-1,-12,-3s-194,-422,-194,-422s-65,47,-65,47z
M834 80h400000v40h-400000z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.08494599999999997em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.93em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3614499999999998em&quot;&gt;&lt;span style=&quot;top:-3.4534200000000004em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen nulldelimiter sizing reset-size3 size6&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2971857142857144em&quot;&gt;&lt;span style=&quot;top:-2.5061857142857145em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9384399999999999em&quot;&gt;&lt;span style=&quot;top:-2.93844em;margin-right:0.1em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.64444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2255000000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line mtight&quot; style=&quot;border-bottom-width:0.049em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.5483000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3448em;margin-left:0em;margin-right:0.1em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.65952em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31472em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;μ&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.04844em&quot;&gt;&lt;span style=&quot;top:-3.04844em;margin-right:0.1em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.64444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.49381428571428565em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter sizing reset-size3 size6&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;정규분포의 확률밀도함수들은 두가지 항으로 이루어져있다.&lt;ol&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;msqrt&gt;&lt;mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;π&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/msqrt&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{1}{\sqrt{2\pi\sigma^2}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.383108em;vertical-align:-0.5379999999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.845108em&quot;&gt;&lt;span style=&quot;top:-2.5153525em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord sqrt mtight&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9637821428571429em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;padding-left:0.833em&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;π&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7463142857142857em&quot;&gt;&lt;span style=&quot;top:-2.786em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.923782142857143em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail mtight&quot; style=&quot;min-width:0.853em;height:1.08em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;1.08em&quot; viewBox=&quot;0 0 400000 1080&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M95,702
c-2.7,0,-7.17,-2.7,-13.5,-8c-5.8,-5.3,-9.5,-10,-9.5,-14
c0,-2,0.3,-3.3,1,-4c1.3,-2.7,23.83,-20.7,67.5,-54
c44.2,-33.3,65.8,-50.3,66.5,-51c1.3,-1.3,3,-2,5,-2c4.7,0,8.7,3.3,12,10
s173,378,173,378c0.7,0,35.3,-71,104,-213c68.7,-142,137.5,-285,206.5,-429
c69,-144,104.5,-217.7,106.5,-221
l0 -0
c5.3,-9.3,12,-14,20,-14
H400000v40H845.2724
s-225.272,467,-225.272,467s-235,486,-235,486c-2.7,4.7,-9,7,-19,7
c-6,0,-10,-1,-12,-3s-194,-422,-194,-422s-65,47,-65,47z
M834 80h400000v40h-400000z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.07621785714285711em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.394em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.5379999999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mrow&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;e^{-\frac{\vert x_i-\mu\vert^2}{2\sigma^2}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.3614499999999998em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3614499999999998em&quot;&gt;&lt;span style=&quot;top:-3.4534200000000004em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen nulldelimiter sizing reset-size3 size6&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2971857142857144em&quot;&gt;&lt;span style=&quot;top:-2.5061857142857145em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9384399999999999em&quot;&gt;&lt;span style=&quot;top:-2.93844em;margin-right:0.1em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.64444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2255000000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line mtight&quot; style=&quot;border-bottom-width:0.049em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.5483000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3448em;margin-left:0em;margin-right:0.1em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.65952em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31472em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;μ&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.04844em&quot;&gt;&lt;span style=&quot;top:-3.04844em;margin-right:0.1em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.64444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.49381428571428565em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter sizing reset-size3 size6&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;/ol&gt;&lt;/li&gt;&lt;li&gt;두 항의 곱을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 연산으로 취하므로, 덧셈 형태로 구할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mfrac&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mfrac&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;π&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;= -\frac{n}{2}\log2\pi\sigma^2 - \displaystyle\sum^n_{i=1}\frac{\vert x_i-\mu\vert^2}{2\sigma^2}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.36687em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.7935600000000003em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.10756em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;π&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.929066em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.491108em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.740108em&quot;&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;정리하면 위와 같은 식이 되는데, 왼쪽 수식(A라고 하자)은 모수 중 &lt;code&gt;분산&lt;/code&gt;만 사용되고, 오른쪽 수식(B라고 하자)은 &lt;code&gt;분산&lt;/code&gt;과 &lt;code&gt;평균&lt;/code&gt;이 모두 사용되는 항이다.&lt;/p&gt;&lt;p&gt;이 수식(A,B)들을 가지고 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta = (\mu,\sigma)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해 미분하여 최적화해볼 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;두 미분이 모두 0이되는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mu&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sigma&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/strong&gt; 를 찾으면 가능도를 최대화하게 된다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;/mrow&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mfrac&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mfrac&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;/mfrac&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msup&gt;&lt;/mfrac&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;0 = \frac{\partial\log L}{\partial\mu}= - \displaystyle\sum^n_{i=1}\frac{x_i-\mu}{\sigma^2}\\
0 = \frac{\partial\log L}{\partial\sigma} = -\frac{n}{\sigma} + \frac{1}{\sigma^3}\sum^n_{i=1}\vert x_i-\mu\vert^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.25188em;vertical-align:-0.8804400000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3714399999999998em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8804400000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.929066em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2603300000000002em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.740108em&quot;&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.0574399999999997em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3714399999999998em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.7935600000000003em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.10756em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.929066em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.740108em&quot;&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1141079999999999em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;평균 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mu&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 으로 미분하면, 수식 A는 버려지고 B만 남게된다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;표준편차 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sigma&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 로 미분하면, 수식 A와 B가 위와 같이 정리된다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;각각의 파라미터에 대해 0이 되는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mu&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sigma&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 구하면 가능도를 최대화해주는 모수를 찾을 수 있다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;/mrow&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mfrac&gt;&lt;mtext&gt;  &lt;/mtext&gt;&lt;mo&gt;⇒&lt;/mo&gt;&lt;mtext&gt;  &lt;/mtext&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mrow&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mfrac&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mfrac&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;/mfrac&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;msup&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msup&gt;&lt;/mfrac&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mtext&gt;  &lt;/mtext&gt;&lt;mo&gt;⇒&lt;/mo&gt;&lt;mtext&gt;  &lt;/mtext&gt;&lt;msubsup&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mrow&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;/mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msubsup&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mfrac&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;0 = \frac{\partial\log L}{\partial\mu}= - \displaystyle\sum^n_{i=1}\frac{x_i-\mu}{\sigma^2} \ \ \Rightarrow \ \ \hat{\mu}_{MLE} = \frac{1}{n}\sum^n_{i=1}x_i\\
0 = \frac{\partial\log L}{\partial\sigma} = -\frac{n}{\sigma} + \frac{1}{\sigma^3}\sum^n_{i=1}\vert x_i-\mu\vert^2\ \ \Rightarrow \ \ \hat{\sigma}^2_{MLE} = \frac{1}{n}\sum^n_{i=1}(x_i-\mu)^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.25188em;vertical-align:-0.8804400000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3714399999999998em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8804400000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.929066em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.2603300000000002em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.740108em&quot;&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;⇒&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.22222em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.929066em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.0574399999999997em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3714399999999998em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.7935600000000003em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.10756em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.929066em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.740108em&quot;&gt;&lt;span style=&quot;top:-2.9890000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1141079999999999em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;⇒&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1111079999999998em;vertical-align:-0.247em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4530000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.247em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.929066em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1141079999999999em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;결국 가능도를 최대화하는 추정 평균 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;μ&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mrow&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\hat{\mu}_{MLE}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;μ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.22222em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 데이터들의 산술평균으로, 추정 분산 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mrow&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;/mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\hat{\sigma}^2_{MLE}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.089439em;vertical-align:-0.275331em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.25em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-2.424669em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.275331em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;은 (데이터-평균)&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mrow&gt;&lt;/mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8141079999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 산술평균이 된다.&lt;/li&gt;&lt;li&gt;이 때 주의할 점은, &lt;div&gt;표본분산을 구할때는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N-1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.76666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 나누었는데, &lt;strong&gt;&lt;code&gt;MLE&lt;/code&gt;&lt;/strong&gt;로 구한 분산은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;N&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;으로 나눈다&lt;/div&gt;는 것이다.&lt;ul&gt;&lt;li&gt;&lt;strong&gt;즉, &lt;code&gt;MLE&lt;/code&gt;는 불편추정량을 보장하진 않는다.&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;그러나 통계에서 이야기하는 consistency는 보장한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-9&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-9&quot; aria-label=&quot; 9 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;최대가능도-추정법-예제--카테고리-분포&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%B5%9C%EB%8C%80%EA%B0%80%EB%8A%A5%EB%8F%84-%EC%B6%94%EC%A0%95%EB%B2%95-%EC%98%88%EC%A0%9C--%EC%B9%B4%ED%85%8C%EA%B3%A0%EB%A6%AC-%EB%B6%84%ED%8F%AC&quot; aria-label=&quot;최대가능도 추정법 예제  카테고리 분포 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;최대가능도 추정법 예제 : 카테고리 분포&lt;/h3&gt;&lt;p&gt;&lt;code&gt;카테고리분포&lt;/code&gt; Multinoulli&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;;&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(x;p_1,\dots,p_d)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 따르는 확률변수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로부터 독립적인 표본 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;{&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;}&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\{x_1,\dots,x_n\}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;}&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 얻었을 때, &lt;strong&gt;&lt;code&gt;최대가능도추정법(MLE)&lt;/code&gt;&lt;/strong&gt;을 이용하여 모수를 추정해보자.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;카테고리분포&lt;/code&gt;는 이산확률변수에 해당한다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;베르누이분포&lt;/code&gt;가 0,1 두개의 값만을 골랐다면, &lt;code&gt;카테고리분포&lt;/code&gt;는 선택지를 다차원으로 확장한 개념이다.&lt;ul&gt;&lt;li&gt;이 때, 선택값은 1이고 나머지는 0인 &lt;code&gt;원-핫벡터&lt;/code&gt;로 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;값을 표현한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;code&gt;카테고리분포&lt;/code&gt;에서 사용하는 모수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;‘&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;‘&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;`p_1,\dots,p_d`&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;‘&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;‘&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 는 &lt;strong&gt;1 에서부터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 차원까지 어떤 값이 1 (또는 0)이 될 확률&lt;/strong&gt;을 의미하는 통계치이다. 따라서 &lt;strong&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;부터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_d&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 까지를 모두 더했을 때 1이 나온다(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sum_{k=1}^dp_k=1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.2887179999999998em;vertical-align:-0.29971000000000003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9890079999999999em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2029em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29971000000000003em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)는 특징을 가지고 있다.&lt;/strong&gt;&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mrow&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;arg max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;;&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;arg max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;munderover&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/munderover&gt;&lt;msubsup&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/msubsup&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\hat{\theta}_{MLE} = \underset{p_1,\dots,p_d}{\argmax}\ \log P(x_i;\theta) = \underset{p_1,\dots,p_d}{\argmax} \ \log\bigg(\prod^n_{i=1}\prod^d_{k=1}p_k^{x_{i,k}}\bigg)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1078799999999998em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9578799999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.26344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.16666em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.780548em;vertical-align:-1.030548em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.20556em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31731428571428577em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;minner mtight&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15122857142857138em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.030548em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.1382260000000004em;vertical-align:-1.302113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.20556em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31731428571428577em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;minner mtight&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15122857142857138em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.030548em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∏&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.836113em&quot;&gt;&lt;span style=&quot;top:-1.8478869999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∏&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.300005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.302113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8492719999999999em&quot;&gt;&lt;span style=&quot;top:-2.3986920000000005em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2478800000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29011428571428566em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3013079999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;만약 해당 데이터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_{i,k}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 1이 될 확률이 0이면, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_k^{x_{i,k}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1505799999999997em;vertical-align:-0.3013079999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8492719999999999em&quot;&gt;&lt;span style=&quot;top:-2.3986920000000005em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2478800000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29011428571428566em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3013079999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_k^0&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.097216em;vertical-align:-0.2831079999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-2.4168920000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2831079999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이 되어 1이 된다.&lt;/li&gt;&lt;li&gt;만약 해당 데이터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_{i,k}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 1이 될 확률이 1이면, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/msubsup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_k^{x_{i,k}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1505799999999997em;vertical-align:-0.3013079999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8492719999999999em&quot;&gt;&lt;span style=&quot;top:-2.3986920000000005em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2478800000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29011428571428566em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3013079999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 되어 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 된다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;위의 식을 풀이해보면,&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;munderover&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/munderover&gt;&lt;msubsup&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/msubsup&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/munderover&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log\bigg(\prod^n_{i=1}\prod^d_{k=1}p_k^{x_{i,k}}\bigg) = \sum^d_{k=1}\bigg(\sum^n_{i=1}x_{i,k}\bigg)\log p_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.1382260000000004em;vertical-align:-1.302113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∏&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.836113em&quot;&gt;&lt;span style=&quot;top:-1.8478869999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∏&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.300005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.302113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8492719999999999em&quot;&gt;&lt;span style=&quot;top:-2.3986920000000005em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2478800000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29011428571428566em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3013079999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.1382260000000004em;vertical-align:-1.302113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.836113em&quot;&gt;&lt;span style=&quot;top:-1.8478869999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.300005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.302113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;이 때, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msubsup&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sum_{i=1}^nx_{i,k}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.104002em;vertical-align:-0.29971000000000003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.804292em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2029em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29971000000000003em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 치환할 수 있다.&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는, 주어진 각 데이터&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x_{i,k}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;들에 대해서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;값이 1인 데이터의 개수를 의미한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;치환하여 다시 식을 정리하면,&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;munderover&gt;&lt;mo&gt;∏&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/munderover&gt;&lt;msubsup&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/msubsup&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/munderover&gt;&lt;msub&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mtext&gt;   &lt;/mtext&gt;&lt;mstyle mathsize=&quot;0.7em&quot;&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mstyle mathsize=&quot;1em&quot;&gt;&lt;mtext&gt;   &lt;/mtext&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/munderover&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mstyle&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log\bigg(\prod^n_{i=1}\prod^d_{k=1}p_k^{x_{i,k}}\bigg) = \sum^d_{k=1}n_k\log p_k \ \ \ \scriptsize with\normalsize \ \ \ \sum^d_{k=1}p_k = 1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.1382260000000004em;vertical-align:-1.302113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∏&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.836113em&quot;&gt;&lt;span style=&quot;top:-1.8478869999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∏&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.300005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.302113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8492719999999999em&quot;&gt;&lt;span style=&quot;top:-2.3986920000000005em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2478800000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29011428571428566em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3013079999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.1382260000000004em;vertical-align:-1.302113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.836113em&quot;&gt;&lt;span style=&quot;top:-1.8478869999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.300005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.302113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord mathnormal sizing reset-size6 size3&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;mord mathnormal sizing reset-size6 size3&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mathnormal sizing reset-size6 size3&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal sizing reset-size6 size3&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mspace sizing reset-size6 size6&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace sizing reset-size6 size6&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace sizing reset-size6 size6&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits sizing reset-size6 size6&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.836113em&quot;&gt;&lt;span style=&quot;top:-1.8478869999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.300005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.302113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord sizing reset-size6 size6&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel sizing reset-size6 size6&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord sizing reset-size6 size6&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;오른쪽 제약식을 만족하면서 왼쪽 목적식을 최대화하는 것이 우리가 구하는 &lt;strong&gt;&lt;code&gt;MLE&lt;/code&gt;&lt;/strong&gt;가 된다.&lt;/li&gt;&lt;li&gt;이렇게 목적식에 제약식이 있는 경우에는, 그냥 미분값이 0이 되는 값을 구하는 것이 아니라, &lt;strong&gt;&lt;code&gt;라그랑주 승수법(Lagrange multiplier method)&lt;/code&gt;&lt;/strong&gt;을 이용하여 목적식을 수정해준다.&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;라그랑주 승수법&lt;/code&gt;&lt;/strong&gt;은 최적화 문제를 푸는 데에 사용된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo&gt;⇒&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/munderover&gt;&lt;msub&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/munder&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\Rightarrow \mathcal{L}(p_1,\dots,p_k,\lambda) = \displaystyle\sum^d_{k=1}n_k\log p_k + \lambda(1-\sum_kp_k)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.36687em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;⇒&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;λ&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.1382260000000004em;vertical-align:-1.302113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.836113em&quot;&gt;&lt;span style=&quot;top:-1.8478869999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.300005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.302113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;λ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.3521180000000004em;vertical-align:-1.3021129999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000005em&quot;&gt;&lt;span style=&quot;top:-1.847887em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.3021129999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;라그랑주 승수법&lt;/code&gt;&lt;/strong&gt;을 이용해, 제약식을 양변으로 넘겨준 상태에서 라그랑주 승수에 해당하는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\lambda&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;λ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 를 곱해준 식을 목적식에 더해주어서 새로운 목적식을 만들어 준다.&lt;ul&gt;&lt;li&gt;이 새로운 목적식의 최적화로, 제약식도 만족하면서, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log L&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 만족시키는 모수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_1,\dots ,p_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 구할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;msub&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mfrac&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;λ&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/munderover&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;0 - \frac{\partial\mathcal{L}}{\partial p_k} = \frac{n_k}{p_k} - \lambda \\
0 - \frac{\partial\mathcal{L}}{\partial \lambda} = 1 - \displaystyle\sum^d_{k=1}p_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.25188em;vertical-align:-0.8804400000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8804400000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.988em;vertical-align:-0.8804400000000001em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1075599999999999em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8804400000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;λ&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.05744em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;λ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.72777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.1382260000000004em;vertical-align:-1.302113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.836113em&quot;&gt;&lt;span style=&quot;top:-1.8478869999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.300005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.302113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이를 조합하면, &lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;msub&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mrow&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/munderover&gt;&lt;msub&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_k = \frac{n_k}{\sum^d_{k=1}n_k}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.286278em;vertical-align:-1.178718em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.1075599999999999em&quot;&gt;&lt;span style=&quot;top:-2.120992em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9890079999999999em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2029em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29971000000000003em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.178718em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;분모에 해당하는 값은, 데이터 개수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 과 같다.&lt;/li&gt;&lt;li&gt;그러므로, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;msub&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p_k = \frac{n_k}{n}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.062252em;vertical-align:-0.345em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.717252em&quot;&gt;&lt;span style=&quot;top:-2.6550000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.41586em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15122857142857138em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.345em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이다.&lt;/li&gt;&lt;li&gt;즉, &lt;code&gt;카테고리분포&lt;/code&gt;의 &lt;strong&gt;&lt;code&gt;MLE&lt;/code&gt;&lt;/strong&gt;는 각각의 class에 해당하는 count 수, 즉 &lt;strong&gt;&lt;div&gt;경우의수를 세어서 전체 중의 비율을 구하는 것&lt;/div&gt;&lt;/strong&gt;임을 알 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-10&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-10&quot; aria-label=&quot; 10 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;딥러닝에서-최대가능도-추정법&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%94%A5%EB%9F%AC%EB%8B%9D%EC%97%90%EC%84%9C-%EC%B5%9C%EB%8C%80%EA%B0%80%EB%8A%A5%EB%8F%84-%EC%B6%94%EC%A0%95%EB%B2%95&quot; aria-label=&quot;딥러닝에서 최대가능도 추정법 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;딥러닝에서 최대가능도 추정법&lt;/h3&gt;&lt;p&gt;딥러닝에서도 MLE를 사용해서 모델을 학습할 수 있다.&lt;/p&gt;&lt;p&gt;딥러닝 모델(Multi-Layer Perceptron)의 가중치를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta = (W^{(1)},\dots,W^{(L)})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.138em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 라고 표기하자. 이 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째 층의 가중치 행렬을 의미한다.&lt;/p&gt;&lt;p&gt;분류문제를 예로 들어보자. 분류문제에서 &lt;code&gt;소프트맥스&lt;/code&gt; 확률벡터는 &lt;code&gt;카테고리분포&lt;/code&gt;의 모수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(p_1,\dots,p_k)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 를 모델링하는 데에 사용할 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;&lt;code&gt;원-핫벡터&lt;/code&gt;로 표현한 정답레이블 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y= (y_1,\dots,y_k)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 관찰데이터로 이용해서, 확률분포인 &lt;code&gt;소프트맥스&lt;/code&gt; 벡터의 &lt;strong&gt;&lt;code&gt;로그가능도&lt;/code&gt;&lt;/strong&gt;를 최적화하여, 딥러닝 모델의 모수인 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 학습시킬 수 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;mrow&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;E&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;arg max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mfrac&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/munderover&gt;&lt;munderover&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;/munderover&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;M&lt;/mi&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\hat{\theta}_{MLE} = \underset{\theta}{\argmax}\ \frac{1}{n}\displaystyle\sum^n_{i=1}\sum^K_{k=1}y_{i,k}\log(MLP_\theta(x_i)_k)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1078799999999998em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9578799999999998em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.26344em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.16666em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.32833099999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05764em&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.1304490000000005em;vertical-align:-1.302113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.43055999999999994em&quot;&gt;&lt;span style=&quot;top:-2.153452em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.946548em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.6513970000000002em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.3000050000000005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.8283360000000002em&quot;&gt;&lt;span style=&quot;top:-1.8478869999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-4.300005em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.302113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;M&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.13889em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;K&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k=1,\dots,K&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;K&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 모든 클래스에 대해서&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;i=1,\dots,n&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.65952em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8388800000000001em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; : 모든 데이터에 대해서(트레이닝 데이터 포인트 1, 2, ...)&lt;/li&gt;&lt;li&gt;[MLP의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;번째 예측값(의 로그값)]과, [정답레이블에 해당하는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y_{i,k}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;]를 곱하여, 이를 산술평균낸다.&lt;ul&gt;&lt;li&gt;이 때, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y_{i,k}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;k-1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.77777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개의 오답 0과 1개의 정답 1을 원소로 가지는 원-핫벡터이다.&lt;ul&gt;&lt;li&gt;따라서, 정답레이블(즉, 1)에 해당하는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y_{i,k}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 단 한개만 있으므로, [정답레이블에 해당하는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y_{i,k}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.716668em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3361079999999999em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;]를 곱한다고 표현하는 것이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;결국 산술평균낸 것들 중 최대(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;arg max&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\argmax&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)는 곧, 전체 데이터에 대해서 정답인 레이블의 확률을 최대로 만드는 가중치 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;θ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\theta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;θ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 를 의미한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-11&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-11&quot; aria-label=&quot; 11 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;확률분포의-거리&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%99%95%EB%A5%A0%EB%B6%84%ED%8F%AC%EC%9D%98-%EA%B1%B0%EB%A6%AC&quot; aria-label=&quot;확률분포의 거리 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;확률분포의 거리&lt;/h2&gt;&lt;p&gt;기계학습에서 사용되는 &lt;code&gt;손실함수(cost function)&lt;/code&gt;들은 &lt;strong&gt;기계학습 모델이 학습하는 확률분포와, 데이터에서 관찰되는 확률분포간의 거리&lt;/strong&gt;를 통해 유도할 수 있다.&lt;/p&gt;&lt;p&gt;위에서 살펴보았던 &lt;strong&gt;&lt;code&gt;MLE&lt;/code&gt;&lt;/strong&gt;로 추정하게 되는 많은 모델학습방법론이, 실제로는 이 &lt;strong&gt;&lt;code&gt;확률분포의 거리&lt;/code&gt;를 최적화&lt;/strong&gt;하는 것과 아주 밀접한 관련이 있다.&lt;/p&gt;&lt;p&gt;데이터 공간에 두 개의 확률 분포 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(x), Q(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 있을 경우, &lt;strong&gt;두 확률분포 사이 거리(distance)&lt;/strong&gt;를 구할 때 다음과 같은 함수들을 이용할 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;총변동 거리(Total Variation Distance, TV)&lt;/code&gt;&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;쿨백-라이블러 발산 (Kullback-Leibler Divergence, KL)&lt;/code&gt;&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;바슈타인 거리(Wasserstein Distance)&lt;/code&gt;&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-12&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-12&quot; aria-label=&quot; 12 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;쿨백-라이블러-발산&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%BF%A8%EB%B0%B1-%EB%9D%BC%EC%9D%B4%EB%B8%94%EB%9F%AC-%EB%B0%9C%EC%82%B0&quot; aria-label=&quot;쿨백 라이블러 발산 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;쿨백-라이블러 발산&lt;/h3&gt;&lt;p&gt;위에서 살펴본 함수들 중 &lt;strong&gt;&lt;code&gt;쿨백-라이블러 발산&lt;/code&gt;&lt;/strong&gt;을 더 자세히 살펴보자.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;쿨백 라이블러 발산(KL Divergence)&lt;/code&gt;&lt;/strong&gt;의 정의는 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mtext&gt;이산확률변수&lt;/mtext&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;K&lt;/mi&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mtext&gt;연속확률변수&lt;/mtext&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;K&lt;/mi&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;msub&gt;&lt;mo&gt;∫&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;X&lt;/mi&gt;&lt;/msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;/mstyle&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;이산확률변수:\mathbb{KL}(P\Vert Q) = \displaystyle\sum_{x \in \mathcal{X}} P(x) \log\bigg(\frac{P(x)}{Q(x)}\bigg)\\
연속확률변수 : \mathbb{KL}(P\Vert Q) = \displaystyle\int_{\mathcal{X}} P(x) \log\bigg(\frac{P(x)}{Q(x)}\bigg)\\&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;이&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;산&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;확&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;률&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;변&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;수&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mord mathbb&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.771706em;vertical-align:-1.321706em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.8556639999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathcal mtight&quot; style=&quot;margin-right:0.14643em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.321706em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;연&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;속&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;확&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;률&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;변&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;수&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mord mathbb&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.40003em;vertical-align:-0.95003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;margin-right:0.44445em;position:relative;top:-0.0011249999999999316em&quot;&gt;∫&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:-0.433619em&quot;&gt;&lt;span style=&quot;top:-1.7880500000000001em;margin-left:-0.44445em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathcal mtight&quot; style=&quot;margin-right:0.14643em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9119499999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.936em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이러한 정의를 다음과 같이 두개의 항으로 분해할 수 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;K&lt;/mi&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathbb{KL}(P\Vert Q) = -\mathbb{E}_{x\sim P(x)}[\log Q(x)] + \mathbb{E}_{x\sim P(x)}[\log P(x)]&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;K&lt;/span&gt;&lt;span class=&quot;mord mathbb&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1052em;vertical-align:-0.3551999999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34480000000000005em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3551999999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1052em;vertical-align:-0.3551999999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34480000000000005em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3551999999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log Q(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 기댓값 - &lt;strong&gt;&lt;code&gt;크로스 엔트로피&lt;/code&gt;&lt;/strong&gt; —&amp;gt; 이후 negative를 취한다&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log P(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대한 기댓값 - &lt;strong&gt;&lt;code&gt;엔트로피&lt;/code&gt;&lt;/strong&gt;&lt;/li&gt;&lt;li&gt;총 두개의 엔트로피 함수로 표현할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이제 분류문제를 가정해보자.&lt;/p&gt;&lt;p&gt;만약 분류 문제에서 정답레이블을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, 모델의 예측을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Q&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 라고 두면, 최대가능도 추정법에서 사용되는 손실함수는 &lt;strong&gt;&lt;code&gt;쿨백-라이블러 발산&lt;/code&gt;&lt;/strong&gt;의 &lt;strong&gt;&lt;code&gt;크로스 엔트로피&lt;/code&gt;의 역수(minus term)&lt;/strong&gt;와 같다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;다시 말해서, 로그가능도&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;log&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\log L&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;lo&lt;span style=&quot;margin-right:0.01389em&quot;&gt;g&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 최적화(최대화)시키는 것은 정답레이블 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 모델 예측 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Q&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Q&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;Q&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 사이의 거리, 즉 &lt;strong&gt;&lt;code&gt;쿨백-라이블러 발산&lt;/code&gt;을 최소화&lt;/strong&gt;하는 것과 같다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이 과정이 잘 이해가 되지 않는다면, 다음 글과 영상을 참조하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://theeluwin.postype.com/post/6080524&quot;&gt;왜 크로스 엔트로피를 쓸까?&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://3months.tistory.com/436&quot;&gt;Cross-entropy 의 이해: 정보이론과의 관계&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://youtu.be/jMU9G5WEtBc?t=215&quot;&gt;ML lec 6-2: Softmax classifier 의 cost함수&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[인공지능 확률론 기초]]></title><description><![CDATA[확률론 맛보기 by 임성빈 교수님, BoostCamp AI Tech 2주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/08_prob_theory/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/08_prob_theory/</guid><pubDate>Thu, 28 Jan 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 수강한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;인공지능-확률론-기초&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%9D%B8%EA%B3%B5%EC%A7%80%EB%8A%A5-%ED%99%95%EB%A5%A0%EB%A1%A0-%EA%B8%B0%EC%B4%88&quot; aria-label=&quot;인공지능 확률론 기초 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;인공지능 확률론 기초&lt;/h1&gt;&lt;h2 id=&quot;딥러닝에서-확률론이-필요한-이유&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%94%A5%EB%9F%AC%EB%8B%9D%EC%97%90%EC%84%9C-%ED%99%95%EB%A5%A0%EB%A1%A0%EC%9D%B4-%ED%95%84%EC%9A%94%ED%95%9C-%EC%9D%B4%EC%9C%A0&quot; aria-label=&quot;딥러닝에서 확률론이 필요한 이유 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;딥러닝에서 확률론이 필요한 이유&lt;/h2&gt;&lt;p&gt;딥러닝은 &lt;strong&gt;확률론 기반의 기계학습 이론&lt;/strong&gt;에 바탕을 두고 있다.&lt;/p&gt;&lt;p&gt;기계 학습에서 사용되는 &lt;code&gt;손실함수(loss function)&lt;/code&gt;들은 &lt;strong&gt;데이터 공간을 통계적으로 해석해서 유도&lt;/strong&gt;한 것이다. 이런 &lt;strong&gt;손실함수를 가지고 모형들을 학습&lt;/strong&gt;시키므로, 확률론을 이해해야 딥러닝 모형 학습의 원리를 이해할 수 있다.&lt;/p&gt;&lt;p&gt;사례를 들어 살펴보자.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;회귀 분석&lt;/code&gt;&lt;/strong&gt;에서, 손실함수로 사용되는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;‘&lt;/mi&gt;&lt;msub&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;‘&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;`L_2-norm`&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.84444em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;‘&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;‘&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;은 &lt;strong&gt;예측오차의 분산을 최소화&lt;/strong&gt;하는 방향으로 학습하도록 유도한다.&lt;/li&gt;&lt;li&gt;&lt;strong&gt;&lt;code&gt;분류 문제&lt;/code&gt;&lt;/strong&gt;에서, 손실함수로 사용되는 &lt;code&gt;교차엔트로피(cross-entropy)&lt;/code&gt;는 &lt;strong&gt;모델 예측의 불확실성을 최소화&lt;/strong&gt;하는 방향으로 학습하도록 유도한다.&lt;/li&gt;&lt;li&gt;결국, 기계학습에서 사용하는 손실함수는 &lt;strong&gt;데이터 분포와 모델 예측 분포의 차이&lt;/strong&gt;, 즉 &lt;code&gt;예측이 틀릴 위험- risk&lt;/code&gt; 를 최소화하는 방향으로 학습하도록 유도한다. 이 과정에서 &lt;strong&gt;확률론을 기반으로 해석&lt;/strong&gt;한다.&lt;/li&gt;&lt;li&gt;특히, &lt;code&gt;분산&lt;/code&gt; 및 &lt;code&gt;불확실성&lt;/code&gt;을 최소화하려면 측정하는 방법을 알아야한다.&lt;/li&gt;&lt;/ul&gt;&lt;h2 id=&quot;확률분포&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%99%95%EB%A5%A0%EB%B6%84%ED%8F%AC&quot; aria-label=&quot;확률분포 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;확률분포&lt;/h2&gt;&lt;div&gt;&lt;p&gt;이 파트는 데이터가 정답레이블을 항상 가진 지도학습을 상정한다. 만약 정답 레이블이 없다면, 데이터 공간 표기에서 y가 존재하지 않는 그래프를 그려야 한다.&lt;/p&gt;&lt;/div&gt;&lt;ul&gt;&lt;li&gt;데이터 공간을 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;X&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;Y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{X\times Y}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.78333em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.26006em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.2939em&quot;&gt;Y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 표기한다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.09371em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 데이터 공간에서 추출한 확률분포의 표기이다.&lt;ul&gt;&lt;li&gt;데이터의 초상화라고 볼 수있다.&lt;/li&gt;&lt;li&gt;이 때 데이터만 가지고 확률분포 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.09371em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 한번에 아는 것은 불가능하므로, 기계학습을 이용해 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.09371em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 분포를 추론하게 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;데이터는 확률 변수로 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(\bold{x},y) ∼ \mathscr{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.09371em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 라고 표기한다.&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;X&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;Y&lt;/mi&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;(\bold{x},y) \in \mathscr{X\times Y}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.78333em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.26006em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.2939em&quot;&gt;Y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 는 데이터공간상의 관측가능한 데이터에 해당한다.&lt;/li&gt;&lt;li&gt;확률 변수의 종류에 따라 확률 분포를 다르게 모델링한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h3 id=&quot;확률-변수--이산확률변수-vs-연속확률변수&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%99%95%EB%A5%A0-%EB%B3%80%EC%88%98--%EC%9D%B4%EC%82%B0%ED%99%95%EB%A5%A0%EB%B3%80%EC%88%98-vs-%EC%97%B0%EC%86%8D%ED%99%95%EB%A5%A0%EB%B3%80%EC%88%98&quot; aria-label=&quot;확률 변수  이산확률변수 vs 연속확률변수 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;확률 변수 : 이산확률변수 vs 연속확률변수&lt;/h3&gt;&lt;p&gt;데이터 공간 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;X&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;Y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{X\times Y}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.78333em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.26006em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.2939em&quot;&gt;Y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 따라 확률변수가 결정된다고 오해하는 경우가 있는데, 실제로는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.09371em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 의해 결정된다.&lt;/p&gt;&lt;p&gt;확률변수는 확률 분포 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.09371em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 따라 &lt;strong&gt;&lt;code&gt;이산형(discrete)&lt;/code&gt;&lt;/strong&gt;과 &lt;strong&gt;&lt;code&gt;연속형(continuous)&lt;/code&gt;&lt;/strong&gt;로 구분된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;이산형(discrete)&lt;/code&gt;&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;정수 집합, 실수 공간에 속하더라도 연속적이지 않은 변수의 분포(ex- [0.5, 0.5])&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;즉, &lt;strong&gt;실수공간에 있다고 해서 연속형 확률분포는 아니다!&lt;/strong&gt;&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;div&gt;확률 변수가 가질 수 있는 경우의 수를 모두 고려하여 확률을 더해서 모델링&lt;/div&gt;&lt;/strong&gt;한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathbb{P}(X\in A) = \sum_{x \in A} P(X = \bold{x})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;P&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.3717110000000003em;vertical-align:-1.321706em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.8556639999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;A&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.321706em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(X = \bold{x})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 확률변수가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\bold{x}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.44444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;값을 가질 확률이다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;이러한 함수를 &lt;strong&gt;&lt;code&gt;확률질량함수&lt;/code&gt;&lt;/strong&gt;라고 부른다.&lt;/p&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;연속형(continuous)&lt;/code&gt;&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;&lt;strong&gt;&lt;div&gt;데이터 공간에 정의된 확률변수의 밀도(density)위에서의 적분&lt;/div&gt;&lt;/strong&gt;을 통해 모델링한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mo&gt;∫&lt;/mo&gt;&lt;mi&gt;A&lt;/mi&gt;&lt;/msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathbb{P}(X\in A) = \int_A P( \bold{x})d\bold{x}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;P&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;A&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.27195em;vertical-align:-0.9119499999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;margin-right:0.44445em;position:relative;top:-0.0011249999999999316em&quot;&gt;∫&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:-0.433619em&quot;&gt;&lt;span style=&quot;top:-1.7880500000000001em;margin-left:-0.44445em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;A&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9119499999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;이산형 확률변수와 달리, 특정 확률변수가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\bold{x}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.44444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;값을 가질 확률을 구하는 것이 불가능하다.&lt;/p&gt;&lt;/li&gt;&lt;li&gt;&lt;p&gt;그 대신에, 확률변수의 &lt;strong&gt;&lt;code&gt;밀도&lt;/code&gt;&lt;/strong&gt;를 사용한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi&gt;lim&lt;/mi&gt;&lt;mo&gt;⁡&lt;/mo&gt;&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mo&gt;→&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mo&gt;≤&lt;/mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;mo&gt;≤&lt;/mo&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;h&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\bold{x}) = \lim_{h\rarr0}\frac{\mathbb{P}(\bold{x}-h \le X \le \bold{x}+h) }{2h}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.179108em;vertical-align:-0.7521079999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-2.347892em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;→&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;0&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;lim&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7521079999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;P&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≤&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≤&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;h&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;위의 함수를 &lt;strong&gt;&lt;code&gt;밀도함수&lt;/code&gt;&lt;/strong&gt;라고 부르며, 확률이 아닌 &lt;strong&gt;누적확률분포의 변화율&lt;/strong&gt;로 생각해야한다.&lt;/li&gt;&lt;li&gt;따라서 매번 적분을 해야 분포값을 계산할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;보통 컴퓨터에서 사용되는 확률변수는 이산형이지만, 기계학습의 경우 여러 종류의 분포가 있으므로 연속형 호확률변수도 같이 사용하게 된다.&lt;/p&gt;&lt;div&gt;&lt;p&gt;확률변수에는 이산형과 연속형만 있는것이 아니다. 어떤 경우에는 이산형, 어떤경우에는 연속형을 취하는 확률변수도 존재한다.&lt;/p&gt;&lt;/div&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;결합확률분포&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B2%B0%ED%95%A9%ED%99%95%EB%A5%A0%EB%B6%84%ED%8F%AC&quot; aria-label=&quot;결합확률분포 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;결합확률분포&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:836px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:75%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;prob_distribution&quot; title=&quot;prob_distribution&quot; src=&quot;/static/3febf0f101a7b541a28a0250865fb8c6/27b8e/prob_distribution.png&quot; srcSet=&quot;/static/3febf0f101a7b541a28a0250865fb8c6/6f3f2/prob_distribution.png 256w,/static/3febf0f101a7b541a28a0250865fb8c6/01e7c/prob_distribution.png 512w,/static/3febf0f101a7b541a28a0250865fb8c6/27b8e/prob_distribution.png 836w&quot; sizes=&quot;(max-width: 836px) 100vw, 836px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;전체 데이터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\bold{x},y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.63888em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 주어진 상태에서 산정하는 분포를 &lt;strong&gt;&lt;code&gt;결합확률분포(joint probability distribution)&lt;/code&gt;&lt;/strong&gt;라고 하며, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\bold{x},y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 표기한다.&lt;/p&gt;&lt;p&gt;위의 그림에서 각각의 파란 점들은 언뜻보면 연속확률분포처럼 보이지만, 빨간색 격자를 기준으로 위치를 나누어 &lt;strong&gt;이산확률분포로 계산&lt;/strong&gt;할 수 있다. 각각의 칸에 대해서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y=1&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;인 파란색 점들의 개수를 카운팅하면, 주어진 데이터의 결합확률분포를 가지고 원래 확률분포 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.09371em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 모델링할 수 있다.&lt;/p&gt;&lt;p&gt;이 때, 원래 확률 분포 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.09371em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 이산형인지, 연속형인지에 따라 결합확률분포가 결정되는것은 아니다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.09371em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 연속형이어도, 결합확률분포는 이산형일 수 있다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.09371em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 이산형이어도, 결합확률분포는 연속형일 수 있다.&lt;/li&gt;&lt;li&gt;이것은 모델링 방법에 따라 결정되는 문제이므로, &lt;strong&gt;&lt;div&gt;원래 데이터의 확률분포 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;D&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{D}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.09371em&quot;&gt;D&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 주어진 데이터에서 실증적으로 추측한 분포는 다를수도 있다 .&lt;/div&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;컴퓨터를 가지고 추측하기 때문에, 원래 확률분포와 다르더라도 근사할 수 있는 방법을 알 수 있다.&lt;/li&gt;&lt;li&gt;따라서 결합확률분포 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\bold{x},y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 는 주어진 데이터의 모양을 보고 적절하게 선택할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;결합확률&lt;/code&gt;&lt;/strong&gt;은 확률론적으로 &lt;strong&gt;n개의 사건이 동시에 발생할 확률&lt;/strong&gt;을 일컫는다.&lt;/p&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;주변확률분포marginal-probability-distribution&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A3%BC%EB%B3%80%ED%99%95%EB%A5%A0%EB%B6%84%ED%8F%ACmarginal-probability-distribution&quot; aria-label=&quot;주변확률분포marginal probability distribution permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;주변확률분포(Marginal Probability Distribution)&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:722px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:148.046875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;marginal_distribution&quot; title=&quot;marginal_distribution&quot; src=&quot;/static/05559c599f7957a466bbc002a90c90ff/d44c9/marginal_distribution.png&quot; srcSet=&quot;/static/05559c599f7957a466bbc002a90c90ff/6f3f2/marginal_distribution.png 256w,/static/05559c599f7957a466bbc002a90c90ff/01e7c/marginal_distribution.png 512w,/static/05559c599f7957a466bbc002a90c90ff/d44c9/marginal_distribution.png 722w&quot; sizes=&quot;(max-width: 722px) 100vw, 722px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mtext&gt;이산&lt;/mtext&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/munder&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mtext&gt;연속&lt;/mtext&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mo&gt;∫&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/msub&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;d&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;이산 : P(\mathrm{x}) = \sum_yP(\mathrm{x,y})\\
연속 : P(\mathrm{x}) = \int_yP(\mathrm{x,y})\mathrm{dy}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;이&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;산&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.4361180000000004em;vertical-align:-1.386113em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000003em&quot;&gt;&lt;span style=&quot;top:-1.8999949999999999em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.386113em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;연&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;속&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.408058em;vertical-align:-1.048058em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;margin-right:0.44445em;position:relative;top:-0.0011249999999999316em&quot;&gt;∫&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:-0.6105579999999999em&quot;&gt;&lt;span style=&quot;top:-1.7880500000000001em;margin-left:-0.44445em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.048058em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;결합확률분포 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;x&lt;/mi&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\bold{x},y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 각각의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해서&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이산 결합확률분포라면, 모두 더해준다&lt;/li&gt;&lt;li&gt;연속 결합확률분포라면, 적분을 해준다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;이를 통해 &lt;strong&gt;&lt;code&gt;주변확률분포(marginal probability distribution)&lt;/code&gt;&lt;/strong&gt;을 구할 수 있다.&lt;/p&gt;&lt;p&gt;주변확률 분포는 &lt;strong&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대한 정보를 줄 뿐, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대한 정보를 주지는 않는다.&lt;/strong&gt; 위의 그림을 확인하면, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 1이든 2이든 상관없이 점들을 세서 빈도를 계산했을 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 따른 주변확률분포를 구할 수 있다.&lt;/p&gt;&lt;p&gt;확률론적으로 ,&lt;strong&gt;&lt;code&gt;주변확률&lt;/code&gt;&lt;/strong&gt;은  &lt;strong&gt;&lt;code&gt;결합확률&lt;/code&gt;&lt;/strong&gt; 과 대비되는 개념이다. 다른 사건과 결합하지 않은 &lt;strong&gt;개별 사건의 확률&lt;/strong&gt;을 주변확률이라고 부른다.&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;조건부확률분포&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A1%B0%EA%B1%B4%EB%B6%80%ED%99%95%EB%A5%A0%EB%B6%84%ED%8F%AC&quot; aria-label=&quot;조건부확률분포 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;조건부확률분포&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:724px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:147.65625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;conditional_prob_distribution&quot; title=&quot;conditional_prob_distribution&quot; src=&quot;/static/ba2171d9f2a6cc7d9b476643392da61f/a242d/conditional_prob_distribution.png&quot; srcSet=&quot;/static/ba2171d9f2a6cc7d9b476643392da61f/6f3f2/conditional_prob_distribution.png 256w,/static/ba2171d9f2a6cc7d9b476643392da61f/01e7c/conditional_prob_distribution.png 512w,/static/ba2171d9f2a6cc7d9b476643392da61f/a242d/conditional_prob_distribution.png 724w&quot; sizes=&quot;(max-width: 724px) 100vw, 724px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 값과 관계없이 구했던 &lt;code&gt;주변확률분포&lt;/code&gt;와 다르게, &lt;strong&gt;&lt;code&gt;조건부확률분포&lt;/code&gt;&lt;/strong&gt; &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\rm x|y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 특정 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 해당하는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;값을 의미한다.&lt;/p&gt;&lt;p&gt;위의 그림은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 1이라고 주어졌을 때의 조건부확률분포이다. 조건부확률분포 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\rm x|y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;strong&gt;특정 클래스가 주어진 조건에서 데이터의 확률 분포&lt;/strong&gt;를 보여준다. 따라서 &lt;strong&gt;입력 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 출력 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 사이의 관계를 모델링(또는 예측)할 때 사용&lt;/strong&gt;된다.&lt;/p&gt;&lt;p&gt;확률론에서, &lt;strong&gt;&lt;code&gt;조건부확률&lt;/code&gt;&lt;/strong&gt;은 &lt;strong&gt;사건 A가 발생한 경우의 사건 B가 발생할 확률&lt;/strong&gt;을 일컫는다.&lt;/p&gt;&lt;p&gt;위의 결합확률분포, 주변확률분포, 조건부확률분포 개념은 조금 어렵게 설명되어 있는데, 좀 더 쉬운 설명으로는 다음 링크를 참조하자.&lt;/p&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;a href=&quot;https://datascienceschool.net/02%20mathematics/06.05%20%EA%B2%B0%ED%95%A9%ED%99%95%EB%A5%A0%EA%B3%BC%20%EC%A1%B0%EA%B1%B4%EB%B6%80%ED%99%95%EB%A5%A0.html&quot;&gt;결합확률분포와 조건부확률분포 - 데이터 사이언스 스쿨&lt;/a&gt;&lt;/p&gt;&lt;h3 id=&quot;조건부확률과-기계학습&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%A1%B0%EA%B1%B4%EB%B6%80%ED%99%95%EB%A5%A0%EA%B3%BC-%EA%B8%B0%EA%B3%84%ED%95%99%EC%8A%B5&quot; aria-label=&quot;조건부확률과 기계학습 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;조건부확률과 기계학습&lt;/h3&gt;&lt;p&gt;조건부확률 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\rm y|x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 입력변수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해 정답이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;일 확률을 의미한다. &lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때, 연속확률분포의 경우 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\rm y|x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 확률이 아니라 밀도로 해석된다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;로지스틱 회귀에서 사용했던 [선형모델 + 소프트맥스 함수] 결합은 &lt;strong&gt;데이터에서 추출된 패턴을 기반으로 확률을 해석&lt;/strong&gt;하는 데에 사용된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;분류 문제에서 소프트맥스에 선형모델을 집어넣으면 확률벡터를 얻을 수 있었다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;분류 문제에서 softmax(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W\phi + b)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;ϕ&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 &lt;strong&gt;데이터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로부터 추출된 특징패턴&lt;/strong&gt; &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\phi(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;ϕ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 가중치행렬 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 통해 조건부 확률 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\rm y | x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, 즉 입력값이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;일때 정답이 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\rm y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;일 확률을 계산한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;기호적으로는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\rm y|\phi(x))&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;ϕ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라고 쓰기도 한다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;회귀문제의 경우 보통 &lt;code&gt;연속확률변수&lt;/code&gt;를 다루므로, 확률로 해석하기 어렵고 &lt;strong&gt;밀도로 해석&lt;/strong&gt;해야한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;p&gt;이 경우 조건부 확률 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;y&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(\rm y |x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathrm&quot; style=&quot;margin-right:0.01389em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathrm&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;가 아닌 &lt;strong&gt;&lt;code&gt;조건부기대값&lt;/code&gt;&lt;/strong&gt; &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathbb{E}[y|x]&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 추정하며, (당연히) 적분으로 표현한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mo&gt;∫&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathbb{E}_{y\sim P(y|x)}[y|x] = \int_yyP(y|x)dy&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1052em;vertical-align:-0.3551999999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34480000000000005em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3551999999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.408058em;vertical-align:-1.048058em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;margin-right:0.44445em;position:relative;top:-0.0011249999999999316em&quot;&gt;∫&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:-0.6105579999999999em&quot;&gt;&lt;span style=&quot;top:-1.7880500000000001em;margin-left:-0.44445em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.048058em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;그렇다면 왜 회귀문제에서 조건부기대값을 사용할까? 이유는 다음과 같다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;a href=&quot;https://blogik.netlify.app/BoostCamp/U_stage/gradient_descent/#%ED%92%80%EC%9D%B4&quot;&gt;회귀 문제를 다룰 때 사용하는 손실함수가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L_2-norm&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 기댓값이었다.&lt;/a&gt;&lt;ul&gt;&lt;li&gt;조건부기대값은 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L_2-norm&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;, 즉  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathbb{E}\Vert y-f(x)\Vert_2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 최소화하는 함수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 일치한다는 것이 수학적으로 증명되어있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;예측 오차의 분산(&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msubsup&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msubsup&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L_2 -Loss = \sum^n_{i=1}(y_i-f(x_i))^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.83333em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.104002em;vertical-align:-0.29971000000000003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol small-op&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.804292em&quot;&gt;&lt;span style=&quot;top:-2.40029em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.2029em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.29971000000000003em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.03588em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.064108em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31166399999999994em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8141079999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;)을 최소화&lt;/strong&gt;하는 적절한 통계치로 사용할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;물론, &lt;strong&gt;조건부기대값이 아닌 다른 수치를 예측모델에 사용할 수도 있다&lt;/strong&gt;.&lt;/p&gt;&lt;p&gt;예를 들어 일반적인 예측보다 좀 더 견실(robust)하게 예측하는 경우에는 조건부기대값보다는 &lt;strong&gt;중앙값(median)을 사용해서 추정&lt;/strong&gt;하기도 한다.&lt;/p&gt;&lt;p&gt;즉, 통계적 모형에서, 원하는 목적에 따라 사용되는 estimator(추정량)이 달라질 수 있다.&lt;/p&gt;&lt;p&gt;어찌됐든,  딥러닝은 다층신경망을 사용하여 &lt;strong&gt;데이터로부터 특징패턴 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;ϕ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\phi&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;ϕ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 를 추출&lt;/strong&gt;한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때, 특징패턴을 학습하기 위해 어떤 &lt;code&gt;손실함수&lt;/code&gt;를 사용할지는 &lt;strong&gt;기계학습문제와 모델에 의해 결정&lt;/strong&gt;된다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;기대값&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EA%B8%B0%EB%8C%80%EA%B0%92&quot; aria-label=&quot;기대값 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;기대값&lt;/h3&gt;&lt;p&gt;확률분포가 주어지면, 데이터를 분석하는 데에 사용가능한 여러 종류의 &lt;strong&gt;&lt;code&gt;통계적 범함수(statistical functional)&lt;/code&gt;&lt;/strong&gt;를 계산할 수 있다.&lt;/p&gt;&lt;p&gt;이 때, &lt;strong&gt;&lt;code&gt;기대값(expectation)&lt;/code&gt;&lt;/strong&gt;은 &lt;strong&gt;데이터를 대표하는 통계량&lt;/strong&gt;이면서 동시에 확률분포를 통해 &lt;strong&gt;다른 통계적 범함수를 계산&lt;/strong&gt;하는 데에 사용된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;기댓값과 평균은 같은 용어로 많이들 사용하고 있는데, 기계학습에서는 좀 더 폭넓은 의미로 사용한다.&lt;/li&gt;&lt;/ul&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mtext&gt;연속확률분포 기대값&lt;/mtext&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mo&gt;∫&lt;/mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/msub&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mtext&gt;이산확률분포 기대값&lt;/mtext&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;x&lt;/mi&gt;&lt;mo&gt;∈&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;연속확률분포\ 기대값 : \mathbb{E}_{\rm x\sim P(x)}[f(x)] = \int_X f(x)p(x)dx\\
이산확률분포\ 기대값 : \mathbb{E}_{\rm x\sim P(x)}[f(x)] = \sum_{\rm x \in X}f(x)P(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;연&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;속&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;확&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;률&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;분&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;포&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;기&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;대&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;값&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1052em;vertical-align:-0.3551999999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34480000000000005em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathrm mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mathrm mtight&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathrm mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3551999999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.27195em;vertical-align:-0.9119499999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;margin-right:0.44445em;position:relative;top:-0.0011249999999999316em&quot;&gt;∫&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:-0.433619em&quot;&gt;&lt;span style=&quot;top:-1.7880500000000001em;margin-left:-0.44445em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9119499999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;이&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;산&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;확&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;률&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;분&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;포&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;기&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;대&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;값&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1052em;vertical-align:-0.3551999999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34480000000000005em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathrm mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mathrm mtight&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathrm mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3551999999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.3717110000000003em;vertical-align:-1.321706em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.050005em&quot;&gt;&lt;span style=&quot;top:-1.8556639999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathrm mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∈&lt;/span&gt;&lt;span class=&quot;mord mathrm mtight&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.0500049999999996em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.321706em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;연속확률분포&lt;/code&gt;&lt;/strong&gt;의 경우 주어진 함수에 &lt;strong&gt;&lt;code&gt;확률밀도함수&lt;/code&gt;&lt;/strong&gt;를 곱한 후 적분한다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;이산확률분포&lt;/code&gt;&lt;/strong&gt;의 경우 주어진 함수에 &lt;strong&gt;&lt;code&gt;확률질량함수&lt;/code&gt;&lt;/strong&gt;를 곱한 다음 급수로 더해준다.&lt;/p&gt;&lt;p&gt;이러한 기대값을 이용해 분산, 첨도, 공분산 등 여러 통계량을 계산할 수 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mtext&gt;분산&lt;/mtext&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;V&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mtext&gt;왜도&lt;/mtext&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mi&gt;S&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo fence=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mo fence=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;msqrt&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;V&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msqrt&gt;&lt;/mfrac&gt;&lt;msup&gt;&lt;mo fence=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;3&lt;/mn&gt;&lt;/msup&gt;&lt;mo fence=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mtext&gt;상관계수&lt;/mtext&gt;&lt;mo&gt;:&lt;/mo&gt;&lt;mi&gt;C&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;v&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;분산:\mathbb{V}(x) = \mathbb{E}_{x\sim P(x)}[(x - \mathbb{E}[x])^2]\\
왜도 : Skewness(x) = \mathbb{E} \Bigg[\Bigg(\frac{x-\mathbb{E}[x]}{\sqrt{\mathbb{V}(x)}}\Bigg)^3\Bigg]\\
상관계수 : Cov(x_1,x_2) = \mathbb{E}_{x_1,x_2\sim P(x_1,x_2)}[x_1-\mathbb{E}[x_1])(x_2-\mathbb{E}[x_2])]&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;분&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;산&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1052em;vertical-align:-0.3551999999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34480000000000005em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3551999999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1141079999999999em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;왜&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;도&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05764em&quot;&gt;S&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:3.2040379999999997em;vertical-align:-1.25003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;[&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.427em&quot;&gt;&lt;span style=&quot;top:-2.175em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord sqrt&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.935em&quot;&gt;&lt;span class=&quot;svg-align&quot; style=&quot;top:-3.2em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.2em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot; style=&quot;padding-left:1em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;V&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.8950000000000005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.2em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;hide-tail&quot; style=&quot;min-width:1.02em;height:1.28em&quot;&gt;&lt;svg width=&quot;400em&quot; height=&quot;1.28em&quot; viewBox=&quot;0 0 400000 1296&quot; preserveAspectRatio=&quot;xMinYMin slice&quot;&gt;&lt;path d=&quot;M263,681c0.7,0,18,39.7,52,119
c34,79.3,68.167,158.7,102.5,238c34.3,79.3,51.8,119.3,52.5,120
c340,-704.7,510.7,-1060.3,512,-1067
l0 -0
c4.7,-7.3,11,-11,19,-11
H40000v40H1012.3
s-271.3,567,-271.3,567c-38.7,80.7,-84,175,-136,283c-52,108,-89.167,185.3,-111.5,232
c-22.3,46.7,-33.8,70.3,-34.5,71c-4.7,4.7,-12.3,7,-23,7s-12,-1,-12,-1
s-109,-253,-109,-253c-72.7,-168,-109.3,-252,-110,-252c-10.7,8,-22,16.7,-34,26
c-22,17.3,-33.3,26,-34,26s-26,-26,-26,-26s76,-59,76,-59s76,-60,76,-60z
M1001 80h400000v40h-400000z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30499999999999994em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.13em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.954008em&quot;&gt;&lt;span style=&quot;top:-4.2029000000000005em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;3&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;delimsizing size4&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;상&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;관&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;계&lt;/span&gt;&lt;span class=&quot;mord hangul_fallback&quot;&gt;수&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;C&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;v&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1052em;vertical-align:-0.3551999999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34480000000000005em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31731428571428577em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31731428571428577em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31731428571428577em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct mtight&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31731428571428577em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3551999999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;위의 기대값 공식에서, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 대신 분산, 왜도, 상관계수등의 통계량에 해당하는 함수들을 집어넣으면 확률분포에서의 통계적 범함수들을 계산할 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 통계값들은, &lt;code&gt;Pandas&lt;/code&gt; 라이브러리에서 &lt;code&gt;df.describe()&lt;/code&gt; 를 사용하면 나오는 것들이다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;결합확률분포를 계산하는 경우에는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;  대신 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;P(x_1,x_2)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;등을 집어넣으면 된다.&lt;/p&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;몬테카를로-샘플링&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%AA%AC%ED%85%8C%EC%B9%B4%EB%A5%BC%EB%A1%9C-%EC%83%98%ED%94%8C%EB%A7%81&quot; aria-label=&quot;몬테카를로 샘플링 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;몬테카를로 샘플링&lt;/h3&gt;&lt;p&gt;만약 확률분포를 잘 알고 있어서, 확률밀도함수를 사용할지 확률질량함수를 사용할지 등을 잘 알고있으면 기계학습에서의 기댓값을 계산하는데에 이용할 수 있을 것이다. 그러나 기계학습의 많은 문제들은 &lt;strong&gt;확률 분포를 명시적으로 모를때가 대부분&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;확률 분포를 모를 때&lt;/strong&gt; 데이터를 이용해서 기대값을 계산하려면, &lt;strong&gt;&lt;code&gt;몬테카를로 샘플링&lt;/code&gt;&lt;/strong&gt; 방법을 사용해야한다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;mo&gt;≈&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mfrac&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/munder&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mfrac&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mtext&gt;    &lt;/mtext&gt;&lt;msup&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo&gt;&lt;mover&gt;&lt;mo&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;.&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;.&lt;/mi&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/mrow&gt;&lt;/mover&gt;&lt;/mo&gt;&lt;mi&gt;P&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathbb{E}_{x\sim P(x)}[f(x)] \approx \frac{1}{N}\displaystyle\sum_i=\frac{1}{N}f(x^{(i)}, \ \ \ \ x^{(i)}\stackrel{i.i.d}{\sim} P(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1052em;vertical-align:-0.3551999999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.34480000000000005em&quot;&gt;&lt;span style=&quot;top:-2.5198em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3551999999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.599109em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000003em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.00744em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.152978em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;∼&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.5668699999999998em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;P&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;타겟 함수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f(x)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 자리에 샘플링한 데이터를 대입하고, 데이터들에 따라서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f(x^{(i)})&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.138em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 산술평균 값을 구하면, 이 값이 기대값에 근사하게 된다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;몬테카를로는 이산형이든 연속형이든 상관없이 성립한다.&lt;/li&gt;&lt;li&gt;단, &lt;strong&gt;&lt;div&gt;샘플링하는 분포에서 독립적으로 샘플링해주어야만 몬테카를로가 제대로 작동한다.&lt;/div&gt;&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;독립추출이 보장된다면 &lt;strong&gt;&lt;code&gt;대수의 법칙(law of large number)&lt;/code&gt;에 의해 수렴성을 보장&lt;/strong&gt;한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;확률분포를 몰라도 샘플링만 가능하다면 기대값을 계산할 수 있으므로, 기계학습과 통계학 모두에서 굉장히 많이 사용되는 도구이다.&lt;/p&gt;&lt;p&gt;몬테카를로법에 대한 위의 수식은 조금 딱딱한 감이 있으므로, 쉽게 이해하기 위해서 다음 글을 읽어보자.
&lt;a href=&quot;https://m.blog.naver.com/msnayana/220372443583&quot;&gt;몬테카를로 방법으로 원주율 구하기&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;몬테카를로-예제--적분-계산하기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%AA%AC%ED%85%8C%EC%B9%B4%EB%A5%BC%EB%A1%9C-%EC%98%88%EC%A0%9C--%EC%A0%81%EB%B6%84-%EA%B3%84%EC%82%B0%ED%95%98%EA%B8%B0&quot; aria-label=&quot;몬테카를로 예제  적분 계산하기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;몬테카를로 예제 : 적분 계산하기&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:934px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:42.578125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAJCAIAAAC9o5sfAAAACXBIWXMAABYlAAAWJQFJUiTwAAABIUlEQVQoz2VR2ZKDIBD0//9tX7IvSTRqvGsxCkoCyOG2UMZs7RQ11Qw0PT1E6x7OZ63N8yW1ce+K24GUUim1fkR0kN06TOJWj+ecpDXr6cs5/4TPQog0TcG31jLGtNYHGTfansfleKvptXxkDUuqsSGzMZZL/TMtTUeqsliWpes6SunpdCKERIHZ9HNSjnnLoHwt+rydzvfx60K+0wflStt1GIYsy6BvjAGFcw4LkXWuJnNSUTAhmNY0Lh7AlwJdbNuWzLgNTTBDt8HI1vbAROw1wXyTAVDJfcYp/IehfAxoi2icFXwGJpZvm/hX9kpFh0l+Ch7TppOAQtFNEEHeQEMDuPsKTtHd++f+kJVa8LeYqhByA9aGrfYVACwY/i+L+AXK9AGLRA/q2wAAAABJRU5ErkJggg==&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;monte_ex&quot; title=&quot;monte_ex&quot; src=&quot;/static/31c7be0defc9c2d1570962a513d99166/078fe/monte_ex.png&quot; srcSet=&quot;/static/31c7be0defc9c2d1570962a513d99166/6f3f2/monte_ex.png 256w,/static/31c7be0defc9c2d1570962a513d99166/01e7c/monte_ex.png 512w,/static/31c7be0defc9c2d1570962a513d99166/078fe/monte_ex.png 934w&quot; sizes=&quot;(max-width: 934px) 100vw, 934px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위와 같은 함수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mrow&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;f(x) = e^{-x^2}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.9869199999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9869199999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913142857142857em&quot;&gt;&lt;span style=&quot;top:-2.931em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;[-1, 1]&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 상에서 적분값을 어떻게 구해보자.&lt;/p&gt;&lt;p&gt;적분 구간이 -1에서 1까지인데, 확률분포가 아닌 공간에서의 적분을 어떻게 할까?&lt;/p&gt;&lt;ul&gt;&lt;li&gt;고등학교 때 배웠던 부정적분의 공식을 통해서 이 함수의 적분을 계산하기는 어렵다.&lt;/li&gt;&lt;/ul&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mfrac&gt;&lt;msubsup&gt;&lt;mo&gt;∫&lt;/mo&gt;&lt;mrow&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msubsup&gt;&lt;msup&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mrow&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;≈&lt;/mo&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi&gt;N&lt;/mi&gt;&lt;/mfrac&gt;&lt;mstyle scriptlevel=&quot;0&quot; displaystyle=&quot;true&quot;&gt;&lt;munder&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;/munder&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mtext&gt;    &lt;/mtext&gt;&lt;msup&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo&gt;∼&lt;/mo&gt;&lt;mi&gt;U&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mstyle&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{1}{2}\int_{-1}^1e^{-x^2}dx \approx \frac{1}{N}\displaystyle\sum_if(x^{(i)}),\ \ \ \ x^{(i)} \sim U(-1,1)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.5342890000000002em;vertical-align:-0.970281em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mop op-symbol large-op&quot; style=&quot;margin-right:0.44445em;position:relative;top:-0.0011249999999999316em&quot;&gt;∫&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.5640080000000003em&quot;&gt;&lt;span style=&quot;top:-1.7880500000000001em;margin-left:-0.44445em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.8129000000000004em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.970281em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0369199999999998em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913142857142857em&quot;&gt;&lt;span style=&quot;top:-2.931em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;≈&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.599109em;vertical-align:-1.277669em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.32144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;N&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.0500050000000003em&quot;&gt;&lt;span style=&quot;top:-1.872331em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.050005em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3.05em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop op-symbol large-op&quot;&gt;∑&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.277669em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot;&gt; &lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.938em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;i&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;∼&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10903em&quot;&gt;U&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ol&gt;&lt;li&gt;구간 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;[-1, 1]&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 길이는 2이므로 균등분포하여 샘플링한다. 확률분포로 바꾸기 위해 구간을 1씩 나누는 균등분포를 사용한다. 즉, 적분값을 2로 나눈다.&lt;/li&gt;&lt;li&gt;이는 기대값을 계산하는 것과 같다. 따라서 몬테카를로 방법을 사용할 수 있다. 함수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mrow&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;e^{-x^2}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.9869199999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.9869199999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8913142857142857em&quot;&gt;&lt;span style=&quot;top:-2.931em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 균등분포에서 추출한 데이터를 집어넣고, 상수평균을 구해준다. 이 값이 원하는 적분값의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;/&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;1/2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;/&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 근사한 값이다.&lt;/li&gt;&lt;li&gt;이제, 마지막으로 양변에 2를 곱해주어 원하는 적분값(에 근사한 값)을 구할 수 있다.&lt;/li&gt;&lt;/ol&gt;&lt;p&gt;이를 파이썬 코드로 옮기면 다음과 같다.&lt;/p&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;bash&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-bash&quot;&gt;BASH&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-bash&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token function&quot;&gt;import&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; numpy as np&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;def mc_int&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;fun, low, high, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;sample_size&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;100&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;repeat&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;10&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;:&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;    int_len &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; np.abs&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;high - low&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;    &lt;/span&gt;&lt;span class=&quot;token function&quot;&gt;stat&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;    &lt;/span&gt;&lt;span class=&quot;token keyword&quot;&gt;for&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token for-or-select variable&quot;&gt;_&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token keyword&quot;&gt;in&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; range&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;repeat&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;:&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;        x &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; np.random.uniform&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;low&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;low, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;high&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;high, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;size&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;sample_size&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;        fun_x &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; fun&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;        int_val &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; int_len * np.mean&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;fun_x&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;        stat.append&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;int_val&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;    &lt;/span&gt;&lt;span class=&quot;token builtin class-name&quot;&gt;return&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; np.mean&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;stat&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, np.std&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;stat&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;def f_x&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;:&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;    &lt;/span&gt;&lt;span class=&quot;token builtin class-name&quot;&gt;return&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; np.exp&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;-x**2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;print&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;mc_int&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;f_x, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;low&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;-1, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;high&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;sample_size&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;10000&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;, &lt;/span&gt;&lt;span class=&quot;token assign-left variable&quot;&gt;repeat&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;100&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;))&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# (1.4935845057262795, 0.0036777255555109412)&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;div&gt;&lt;p&gt;만약 샘플사이즈가 적게 되면, 몬테카를로 법칙이라도 오차범위가 커질 수 있으므로, 적절한 샘플링 개수를 확보해야한다.&lt;/p&gt;&lt;/div&gt;</content:encoded></item><item><title><![CDATA[딥러닝의 기초적 이해 - 선형모델부터 역전파에 이르기까지]]></title><description><![CDATA[딥러닝의 학습방법 by 임성빈 교수님, BoostCamp AI Tech 2주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/07_deep_learning_basic/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/07_deep_learning_basic/</guid><pubDate>Wed, 27 Jan 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 수강한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;딥러닝의-학습방법&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%94%A5%EB%9F%AC%EB%8B%9D%EC%9D%98-%ED%95%99%EC%8A%B5%EB%B0%A9%EB%B2%95&quot; aria-label=&quot;딥러닝의 학습방법 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;딥러닝의 학습방법&lt;/h1&gt;&lt;h2 id=&quot;선형모델&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%84%A0%ED%98%95%EB%AA%A8%EB%8D%B8&quot; aria-label=&quot;선형모델 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;선형모델&lt;/h2&gt;&lt;p&gt;지난시간까지, 데이터를 선형모델로 해석하여 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;값과 선형모델 예측값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\hat{y}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.19444em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 차이의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;r&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L2-norm&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.76666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;r&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 기댓값을 최소화하는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;β&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\beta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05278em&quot;&gt;β&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 찾는것이었다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mi&gt;i&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/mo&gt;&lt;mi&gt;β&lt;/mi&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;E&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;−&lt;/mo&gt;&lt;mover accent=&quot;true&quot;&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo&gt;^&lt;/mo&gt;&lt;/mover&gt;&lt;msub&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∥&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;{\underset{\beta}{min}} \mathbb{E}\Vert y- \hat{y}\Vert_2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.638216em;vertical-align:-0.8882159999999999em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.65952em&quot;&gt;&lt;span style=&quot;top:-2.3478920000000003em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.05278em&quot;&gt;β&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;min&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8882159999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;E&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;−&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord accent&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;accent-body&quot; style=&quot;left:-0.19444em&quot;&gt;&lt;span class=&quot;mord&quot;&gt;^&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.19444em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;∥&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;그러나, 이러한 선형 모델은 단순한 선형문제를 푸는데에는 사용할 수 있겠지만, 분류(classification)이나 더 복잡한 패턴의 문제를 제대로 예측하기가 어렵다.&lt;/p&gt;&lt;p&gt;따라서, 이제부터는 비선형모델인 &lt;strong&gt;&lt;code&gt;신경망(Neural Network)&lt;/code&gt;&lt;/strong&gt;를 사용해 보자.&lt;/p&gt;&lt;p&gt;신경망은 비선형모델이지만, &lt;strong&gt;내부적으로는 선형모델들의 결합을 기반&lt;/strong&gt;으로 만들어져있다.&lt;/p&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;신경망의-수식-표현&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%8B%A0%EA%B2%BD%EB%A7%9D%EC%9D%98-%EC%88%98%EC%8B%9D-%ED%91%9C%ED%98%84&quot; aria-label=&quot;신경망의 수식 표현 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;신경망의 수식 표현&lt;/h2&gt;&lt;h3 id=&quot;선형-모델&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%84%A0%ED%98%95-%EB%AA%A8%EB%8D%B8&quot; aria-label=&quot;선형 모델 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;선형 모델&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:35.15625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAABGUlEQVQY02VQTU/DMAzd//9BIG1cQHCZAE1wm2AjWrt2SfPROE2TJnHI6JiQ8MGx7Lzn57fI/2J/7LreQG90bwxYOGejVM+5sNZqMCml+edifnxE8ClEzGmQvGGStpwweZDQMFWB0Qag8GGKGEPOmBGvYCS031S6FVq2b9A8V/XrbrjbwarBhx0sG0aoGpWLwgbtkYFnQl3ARcZobaFQSpnBhikf6HbvbohbVnH15W5J+/G4lU/Ermu/Yfn+02wP3QWMiOUYLoRzDkBLoYQ+UfdO5Lo2L6VQ0IUpFK3WDinGssyel/3eHEKoqnpmaU+nuSm4YpTP7HOnjKwdr9Yuiod1faSUxR8P8RJ5HJ33fpqmIucK/ktUFHwD/OWPe+exRiMAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;linear_model&quot; title=&quot;linear_model&quot; src=&quot;/static/bc1c7695f00a5fab08c333cce4ad036a/2bef9/linear_model.png&quot; srcSet=&quot;/static/bc1c7695f00a5fab08c333cce4ad036a/6f3f2/linear_model.png 256w,/static/bc1c7695f00a5fab08c333cce4ad036a/01e7c/linear_model.png 512w,/static/bc1c7695f00a5fab08c333cce4ad036a/2bef9/linear_model.png 1024w,/static/bc1c7695f00a5fab08c333cce4ad036a/71c1d/linear_model.png 1536w,/static/bc1c7695f00a5fab08c333cce4ad036a/a878e/linear_model.png 2048w,/static/bc1c7695f00a5fab08c333cce4ad036a/73dae/linear_model.png 2122w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;/mo&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/mo&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mo&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;&lt;munder&gt;&lt;mo&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;/mo&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/munder&gt;&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\underset{(n\times p)}{O} = \underset{(n\times d)}{X} \underset{(d\times p)}{W} + 
\underset{(n\times p)}{b}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.64933em;vertical-align:-0.966em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.68333em&quot;&gt;&lt;span style=&quot;top:-2.3089999999999997em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.966em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.64933em;vertical-align:-0.966em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.68333em&quot;&gt;&lt;span style=&quot;top:-2.3089999999999997em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.966em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.68333em&quot;&gt;&lt;span style=&quot;top:-2.3089999999999997em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.966em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.66044em;vertical-align:-0.966em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mop op-limits&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.69444em&quot;&gt;&lt;span style=&quot;top:-2.3089999999999997em;margin-left:0em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;span class=&quot;mbin mtight&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span&gt;&lt;span class=&quot;mop mathnormal&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.966em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;전체 데이터&lt;/strong&gt;가 모인 행렬 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 한 &lt;strong&gt;행벡터&lt;/strong&gt;는 하나의 점으로 표현되는 &lt;strong&gt;데이터 포인트&lt;/strong&gt;이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 데이터를 출력 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 보내주는 가중치 행렬 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 데이터를 &lt;strong&gt;다른 공간으로 보내주는&lt;/strong&gt; 역할을 한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 절편에 해당하는 행벡터를 모든 행에 복제하여 만든 절편 행렬 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;b&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;ul&gt;&lt;li&gt;각 행들은 전부 같은 값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;[&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;]&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;[b_1, b_2, \cdots, b_p]&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:0em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;]&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 가진다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;div&gt;이 때, 출력벡터 차원(열)은 기존의 $X$ 벡터 차원 $d$ 에서 $p$ 로 바뀌게 된다.&lt;/div&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:848px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:46.09375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAJCAIAAAC9o5sfAAAACXBIWXMAABYlAAAWJQFJUiTwAAABKElEQVQoz1VSi5LDIAjM//9n2zTRxPgEFHtr0t5dmckEkV1gcXp9jKsSa2s6fFESbdpf/0x7v+K9v+PT9auth1RyziRNmsZUIj5uA6NgbBd7iCnlcsX/wKqv3YX7Y84McLe7v90fieq4Q6EOfLeeF2Pn58L1uzLuXWQX+LbGh0nLXsxBmy8+yRHZR7a+zFtCz3vg33HeYOJWmyKvcN0DbQe5SGBZXZ5tupsEcNMhB4Zi+bSNmuBD2ZAFpVDcAhnIuLx5ylTRPIYE+zFao1jEZ4E00G0CMqacC4ckqdRw+siGjzhEI3nLNqxDV4mx5EKIT1L7aqzdHASHVM/nspoNtNB2nhe7H+fuLvAYFUckQFokj7alYoyxvY/frq2CF8f+tewh7XgRZ/4PPpAMwbiVjNYAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;d-to-p&quot; title=&quot;d-to-p&quot; src=&quot;/static/5551d2650565a6cdf3aaaa12e86d7a37/d52e5/d-to-p.png&quot; srcSet=&quot;/static/5551d2650565a6cdf3aaaa12e86d7a37/6f3f2/d-to-p.png 256w,/static/5551d2650565a6cdf3aaaa12e86d7a37/01e7c/d-to-p.png 512w,/static/5551d2650565a6cdf3aaaa12e86d7a37/d52e5/d-to-p.png 848w&quot; sizes=&quot;(max-width: 848px) 100vw, 848px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이는 &lt;strong&gt;&lt;div&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 개의 변수로 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 개의 선형 모델을 만들어서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;p&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 개의 잠재변수를 설명하는 모델&lt;/div&gt;&lt;/strong&gt;과 동일한 의미이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;위의 그림에서 &lt;strong&gt;화살표&lt;/strong&gt;는 가중치 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 의미한다. 즉, 화살표의 갯수는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;d&lt;/mi&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;d \times p&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.77777em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;d&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 가 된다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;소프트맥스softmax-연산&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%86%8C%ED%94%84%ED%8A%B8%EB%A7%A5%EC%8A%A4softmax-%EC%97%B0%EC%82%B0&quot; aria-label=&quot;소프트맥스softmax 연산 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;소프트맥스(softmax) 연산&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;소프트맥스&lt;/code&gt;&lt;/strong&gt; 함수는 &lt;strong&gt;모델의 출력을 확률로 해석&lt;/strong&gt;할 수 있게 변환해주는 연산이다.&lt;/p&gt;&lt;p&gt;&lt;strong&gt;분류 문제&lt;/strong&gt;를 풀 때 선형모델과 소프트맥스 함수를 결합하여 예측할 수 있다.&lt;/p&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;s&lt;/mi&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;f&lt;/mi&gt;&lt;mi&gt;t&lt;/mi&gt;&lt;mi&gt;m&lt;/mi&gt;&lt;mi&gt;a&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mrow&gt;&lt;mo fence=&quot;true&quot;&gt;(&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/msubsup&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;msubsup&gt;&lt;mo&gt;∑&lt;/mo&gt;&lt;mrow&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/mrow&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;/msubsup&gt;&lt;mi&gt;e&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi&gt;p&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo fence=&quot;true&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;softmax(o) = \left ( \frac{exp(o_1)}{\sum_{k=1}^{p}exp(o_k)}, \cdots , \frac{exp(o_p)}{\sum_{k=1}^{p}exp(o_k)} \right )&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.10764em&quot;&gt;f&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;t&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;m&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;a&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.80002em;vertical-align:-0.65002em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;&lt;span class=&quot;mopen delimcenter&quot; style=&quot;top:0em&quot;&gt;&lt;span class=&quot;delimsizing size2&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.01em&quot;&gt;&lt;span style=&quot;top:-2.63614em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mop mtight&quot;&gt;&lt;span class=&quot;mop op-symbol small-op mtight&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7912285714285715em&quot;&gt;&lt;span style=&quot;top:-2.1527714285714286em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.9836857142857145em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3472285714285714em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace mtight&quot; style=&quot;margin-right:0.19516666666666668em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15122857142857138em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.485em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.31731428571428577em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.6069199999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.03232em&quot;&gt;&lt;span style=&quot;top:-2.63614em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mop mtight&quot;&gt;&lt;span class=&quot;mop op-symbol small-op mtight&quot; style=&quot;position:relative;top:-0.0000050000000000050004em&quot;&gt;∑&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.7912285714285715em&quot;&gt;&lt;span style=&quot;top:-2.1527714285714286em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;span class=&quot;mrel mtight&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-2.9836857142857145em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3472285714285714em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace mtight&quot; style=&quot;margin-right:0.19516666666666668em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.3448em&quot;&gt;&lt;span style=&quot;top:-2.3487714285714287em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15122857142857138em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.50732em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;e&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;o&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.16454285714285716em&quot;&gt;&lt;span style=&quot;top:-2.357em;margin-left:0em;margin-right:0.07142857142857144em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.5em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size3 size1 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;p&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.2818857142857143em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.6069199999999999em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose delimcenter&quot; style=&quot;top:0em&quot;&gt;&lt;span class=&quot;delimsizing size2&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;ul&gt;&lt;li&gt;선형모델의 출력 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;O&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 의 행벡터를 softmax함수의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;o&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;o&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;o&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 로 집어넣는다.&lt;/li&gt;&lt;li&gt;softmax 함수는 입력을 확률벡터로 변환시켜 출력한다.&lt;ul&gt;&lt;li&gt;주어진 데이터가 특정 클래스에 속할 확률이 얼마인지를 계산한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;Python으로 구현한 코드는 다음과 같다.&lt;/p&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;python&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-python&quot;&gt;PYTHON&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-python&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token keyword&quot;&gt;def&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token function&quot;&gt;softmax&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;vec&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;    &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 수식과 달리, np.max값을 취해주는 과정이 추가되어 있다.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;    &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# softmax는 지수함수를 취하므로, 너무 큰 값이 들어왔을 경우 overflow가 생길 위험이 있다.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;    &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 이를 방지하기 위해 np.max 값을 vector에다 빼 준 뒤 해당 값을 input으로 적용한다.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;    denumerator &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; np&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;exp&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;vec &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;-&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; np&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token builtin&quot;&gt;max&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;vec&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; axis&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;-&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; keepdims&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token boolean&quot;&gt;True&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;    numerator &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; np&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token builtin&quot;&gt;sum&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;denumerator&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; axis&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;-&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; keepdims&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token boolean&quot;&gt;True&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;    val &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; denumerator &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;/&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; numerator&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;    &lt;/span&gt;&lt;span class=&quot;token keyword&quot;&gt;return&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; val&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;vec &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; np&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;array&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;-&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;-&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;10&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;10&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;softmax&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;vec&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;&amp;#x27;&amp;#x27;&amp;#x27;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;array([[2.44728471e-01, 6.65240956e-01, 9.00305732e-02],&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;       [9.00305732e-02, 2.44728471e-01, 6.65240956e-01],&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;       [2.06106005e-09, 4.53978686e-05, 9.99954600e-01]])&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;[1,2,0]이 [0.24, 0.67, 0.09]라는 확률벡터로 바뀌었다.&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;&amp;#x27;&amp;#x27;&amp;#x27;&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;h1 id=&quot;-7&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-7&quot; aria-label=&quot; 7 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;왜 분류문제에 있어서 최종 출력값이 확률값이 되어야하고, 소프트맥스 함수를 사용해야 하는지&lt;/strong&gt;에 대해서는 다음 링크를 참고한다.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://tensorflow.blog/%ED%95%B4%EC%BB%A4%EC%97%90%EA%B2%8C-%EC%A0%84%ED%95%B4%EB%93%A4%EC%9D%80-%EB%A8%B8%EC%8B%A0%EB%9F%AC%EB%8B%9D-3/&quot;&gt;해커에게 전해들은 머신러닝 #3&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-8&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-8&quot; aria-label=&quot; 8 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;단, 소프트맥스 함수는 &lt;div&gt;학습 시에만 사용하고, 추론을 할 때에는 사용하지 않는다.&lt;/div&gt;&lt;/p&gt;&lt;p&gt;추론을 할 때에는 &lt;strong&gt;출력값에서 최댓값을 가진 주소만 1로 출력&lt;/strong&gt;하는 &lt;strong&gt;&lt;code&gt;원-핫(one-hot) 벡터&lt;/code&gt;&lt;/strong&gt;를 사용하여, &lt;strong&gt;주어진 출력 중 최댓값 주소만 가져가는 형태로 구현&lt;/strong&gt;하기 때문이다.&lt;/p&gt;&lt;h1 id=&quot;-9&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-9&quot; aria-label=&quot; 9 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;선형함수에 소프트맥스 함수를 적용시켜 선형 모델을 분류문제에 알맞은 확률 추측 모델로 바꾸었듯이 함수를 합성하여 출력값을 조정할 수 있다면, , &lt;strong&gt;선형 함수에 다른 함수를 합성하면 비선형 문제도 풀 수 있지 않을까?&lt;/strong&gt;&lt;/p&gt;&lt;h1 id=&quot;-10&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-10&quot; aria-label=&quot; 10 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;활성함수activation-function&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%ED%99%9C%EC%84%B1%ED%95%A8%EC%88%98activation-function&quot; aria-label=&quot;활성함수activation function permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;활성함수(activation function)&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:33.984375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAABKElEQVQY01VQy07EMAzs//8NFw5ckbhwRVoQj+52pb6b5h07dlK8cIHRyLJGtuWZ5viPWisAEFHOmZlFoR9wKYgoiujee6mZqHHOcaYD8kEsZMBLdxVxWVattSxrbcZxAsBlWZzzxpi2PW/bppRq0Hg/qdhNdTPpPMR2aCc1rErtuzb2NiLdrqdp3tVunRca60KMtR5NQfImeB3YxbBZsj6G6GIiBASIkBNgSuBvCClGzhiczZAKc8Ol2kTdsPbrHkkM4yGeAd4XGG1Ce133Uxcf5/Lcp6dRj4PlOZXe0eagCTE5H2PKP6GUIOYlpEy98vNugt769eXN3V3woQ33n8PX64gfKp/mdF1NM8+LBMPEfzOXQ/J2RnmdMQMWk6vDYkv9HavCUvgbrguQKVyTHyUAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;activation_function&quot; title=&quot;activation_function&quot; src=&quot;/static/a2a099ba0ccb9195eef849a7f7a66220/2bef9/activation_function.png&quot; srcSet=&quot;/static/a2a099ba0ccb9195eef849a7f7a66220/6f3f2/activation_function.png 256w,/static/a2a099ba0ccb9195eef849a7f7a66220/01e7c/activation_function.png 512w,/static/a2a099ba0ccb9195eef849a7f7a66220/2bef9/activation_function.png 1024w,/static/a2a099ba0ccb9195eef849a7f7a66220/71c1d/activation_function.png 1536w,/static/a2a099ba0ccb9195eef849a7f7a66220/a878e/activation_function.png 2048w,/static/a2a099ba0ccb9195eef849a7f7a66220/b3608/activation_function.png 2160w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-11&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-11&quot; aria-label=&quot; 11 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;활성함수&lt;/code&gt;&lt;/strong&gt;는 실수값을 입력으로 받아서 다시 실수값으로 뱉는 &lt;strong&gt;비선형(nonlinear) 함수&lt;/strong&gt;이다.&lt;/p&gt;&lt;p&gt;이를 &lt;strong&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;double-struck&quot;&gt;R&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathbb{R}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68889em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbb&quot;&gt;R&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 위에 정의된 비선형함수&lt;/strong&gt;라고 하며, 딥러닝에서 매우 중요한 개념이다.&lt;/p&gt;&lt;h1 id=&quot;-12&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-12&quot; aria-label=&quot; 12 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;ul&gt;&lt;li&gt;선형모델을 입력으로 받아서 각각의 원소에 대하여 적용된다.&lt;/li&gt;&lt;li&gt;정확히 말하면, 활성함수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sigma&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;는 비선형 함수로, 입력 &lt;code&gt;잠재벡터&lt;/code&gt; &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;z&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;q&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\bold{z} = (z_1,\cdots,z_q)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.44444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.036108em;vertical-align:-0.286108em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15139200000000003em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03588em&quot;&gt;q&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.286108em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 각 노드에 개별적으로 적용하여 새로운 &lt;code&gt;잠재 벡터&lt;/code&gt; &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;H&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mtext&gt; &lt;/mtext&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;n&lt;/mi&gt;&lt;/msub&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\bold{H} = (\sigma(z_1),\cdots,\sigma(z_n))&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68611em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.30110799999999993em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.151392em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;n&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;을 만든다.&lt;ul&gt;&lt;li&gt;이 잠재벡터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;H&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 &lt;strong&gt;&lt;code&gt;히든(Hidden) 벡터&lt;/code&gt;&lt;/strong&gt; 또는 &lt;strong&gt;&lt;code&gt;뉴런(Neuron)&lt;/code&gt;&lt;/strong&gt;이라 부르기도 한다.&lt;/li&gt;&lt;li&gt;이런 뉴런의 집합체를 &lt;strong&gt;&lt;code&gt;신경망, Neural Network&lt;/code&gt;&lt;/strong&gt;라고 부르는 것이다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;선형 모델에 활성함수를 한차례 씌워 비선형 모델로 전환한 이런 &lt;strong&gt;기본적인 신경망&lt;/strong&gt;을 역사적으로 &lt;strong&gt;&lt;code&gt;퍼셉트론(Perceptron)&lt;/code&gt;&lt;/strong&gt;이라고 부른다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-13&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-13&quot; aria-label=&quot; 13 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;무엇보다 명심해야할 점은, &lt;strong&gt;&lt;div&gt;활성함수를 쓰지 않으면 딥러닝은 선형모델과 차이가 없다&lt;/div&gt;&lt;/strong&gt;는 것이다.&lt;/p&gt;&lt;h1 id=&quot;-14&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-14&quot; aria-label=&quot; 14 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h1 id=&quot;-15&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-15&quot; aria-label=&quot; 15 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:25.390625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAFCAIAAADKYVtkAAAACXBIWXMAABYlAAAWJQFJUiTwAAAApElEQVQY0y1QWxIDIQjb+1+13Z92RcJDGrSMw2hISMarqlaVu4dbhGdY5loR6dYn04EwIyd3rVWNuFN4NUomzACViTl5ISkyAcjznCdVu4fOqSK944jDw2BhTr8WsbmbqimIrF3RyYwekEmInC3uFR1bZI4xAKs6SAyRz/dR4CAUqFI7aXKQi7P3fb9e7zHEuNv/A1ivY0YS1nZof36BM0zbkvsDkQEmWIAvdZAAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;activation_function_graph&quot; title=&quot;activation_function_graph&quot; src=&quot;/static/5767cddb69f5b16073d3d3391b60903a/2bef9/activation_function_graph.png&quot; srcSet=&quot;/static/5767cddb69f5b16073d3d3391b60903a/6f3f2/activation_function_graph.png 256w,/static/5767cddb69f5b16073d3d3391b60903a/01e7c/activation_function_graph.png 512w,/static/5767cddb69f5b16073d3d3391b60903a/2bef9/activation_function_graph.png 1024w,/static/5767cddb69f5b16073d3d3391b60903a/71c1d/activation_function_graph.png 1536w,/static/5767cddb69f5b16073d3d3391b60903a/a878e/activation_function_graph.png 2048w,/static/5767cddb69f5b16073d3d3391b60903a/96191/activation_function_graph.png 2176w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-16&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-16&quot; aria-label=&quot; 16 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;과거에는 활성함수로 &lt;strong&gt;&lt;code&gt;시그모이드(sigmoid)&lt;/code&gt;&lt;/strong&gt; 함수나 &lt;strong&gt;&lt;code&gt;하이퍼볼릭탄젠트(tanh)&lt;/code&gt;&lt;/strong&gt; 함수를 많이 사용되었으며, 70년정도 된 신경망과 퍼셉트론의 역사에서 큰 역할을 했다.&lt;/p&gt;&lt;p&gt;그러나 오늘날 딥러닝에 가장 많이 쓰이는 함수는 &lt;strong&gt;&lt;code&gt;ReLU&lt;/code&gt;&lt;/strong&gt; 함수로, 언뜻 보면 선형함수처럼 보이지만 전형적인 비선형함수이며 비선형함수로서의 좋은 성질들을 많이 가지고 있다.&lt;/p&gt;&lt;p&gt;왜 ReLU 함수가 시그모이드, tanh 함수보다 나은지에 대해서는 다음 링크를 참조하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://stats.stackexchange.com/questions/126238/what-are-the-advantages-of-relu-over-sigmoid-function-in-deep-neural-networks&quot;&gt;What are the advantages of ReLU over sigmoid function in deep neural networks?&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://reniew.github.io/12/&quot;&gt;딥러닝에서 사용하는 활성화함수&lt;/a&gt;&lt;/p&gt;&lt;h1 id=&quot;-17&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-17&quot; aria-label=&quot; 17 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;요약하자면 다음과 같다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;시그모이드 함수와 tanh 함수는, Gradient Vanishing 가능성이 비교적 높다.&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;시그모이드 함수는 미분값(gradient 값)의 최대가 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mn&gt;4&lt;/mn&gt;&lt;/mfrac&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;0&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{1}{4}(x=0)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.190108em;vertical-align:-0.345em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.845108em&quot;&gt;&lt;span style=&quot;top:-2.6550000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;4&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.394em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.345em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 밖에 되지 않고, 어느 정도 input값이 올라가면 거의 0에 수렴한다. 이 경우 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∣&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;|x|&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;∣&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;값이 커질수록 backpropagation 과정에서 미분값이 소실될 가능성이 크다.&lt;/li&gt;&lt;li&gt;tanh 함수도 비슷한 문제가 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;시그모이드 함수는 함수값 중심이 0이 아니다.&lt;/strong&gt;&lt;ul&gt;&lt;li&gt;모든 파라미터의 미분값이 같은 부호를 가지게 되어, 같은 방향으로 update되게 된다.&lt;/li&gt;&lt;li&gt;이는 학습을 지그재그형태로 만들어 학습이 느려지는 원인이 된다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;strong&gt;시그모이드 함수는 exp 함수 사용시 비용이 크다.&lt;/strong&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-18&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-18&quot; aria-label=&quot; 18 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;신경망&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%8B%A0%EA%B2%BD%EB%A7%9D&quot; aria-label=&quot;신경망 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;신경망&lt;/h3&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;신경망&lt;/code&gt;&lt;/strong&gt;은 &lt;strong&gt;&lt;code&gt;선형모델&lt;/code&gt;&lt;/strong&gt;과 &lt;strong&gt;&lt;code&gt;활성함수&lt;/code&gt;&lt;/strong&gt;를 합성한 함수이다.&lt;/p&gt;&lt;p&gt;위에서 짚었던 활성함수 개념을 토대로, 신경망을 층층이 만들어 보자.&lt;/p&gt;&lt;h1 id=&quot;-19&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-19&quot; aria-label=&quot; 19 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:34.765625%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAABI0lEQVQY0z1Ry07DMBDs/38QnPiEXuASiaaChhL8ih0n9j5iJ2wTwcqHWWlnZ3Z82vaqta7/YD3gtu4lgJmttYgUQihFRmqMU875tA9tQIsLkx8nwIV4cdb1fd80DQAopRDxdrsNg++6r7a9eh8ul1YbczpUMrIdovMj8iKL2k65MBrrnDD8YKz5UUob7aXxYYyTNnaa04MsyimT8PctxAAhzjFlhiz2xpjGEDNgSoCAJO4RICciOshronLXQ9frGQsjbYQ54Wcocxoxdiq9KT7rctbxQ0/FQbW5jAkfZBGLc04ZgVjykrPFQgLqPXjv4vB9GV6u8PyBT61+bXp4d9wYurvpRMQSjxxHzEfCR9TMCz8ssuzCZaYa5S0Vt+3vK2r9BTlGkDcRWritAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;2_layer_nn&quot; title=&quot;2_layer_nn&quot; src=&quot;/static/7ee09533ab2a14bd57bf0f72e9441782/2bef9/2_layer_nn.png&quot; srcSet=&quot;/static/7ee09533ab2a14bd57bf0f72e9441782/6f3f2/2_layer_nn.png 256w,/static/7ee09533ab2a14bd57bf0f72e9441782/01e7c/2_layer_nn.png 512w,/static/7ee09533ab2a14bd57bf0f72e9441782/2bef9/2_layer_nn.png 1024w,/static/7ee09533ab2a14bd57bf0f72e9441782/71c1d/2_layer_nn.png 1536w,/static/7ee09533ab2a14bd57bf0f72e9441782/a878e/2_layer_nn.png 2048w,/static/7ee09533ab2a14bd57bf0f72e9441782/ba715/2_layer_nn.png 2210w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-20&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-20&quot; aria-label=&quot; 20 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;2층 신경망(2-layer NN) 구현하기&lt;/strong&gt;&lt;/p&gt;&lt;ol&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 가중치 행렬 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^{(1)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8879999999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;b^{(1)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8879999999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 선형변환시켜 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;Z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;라는 출력을 받는다.&lt;/li&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;Z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;Z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;Z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 활성함수 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;σ&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\sigma&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;σ&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 씌워서 히든벡터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;H&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 만든다. - 이 때 활성함수는 Z의 각 원소&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msub&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mi&gt;k&lt;/mi&gt;&lt;/msub&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z_k&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.58056em;vertical-align:-0.15em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.33610799999999996em&quot;&gt;&lt;span style=&quot;top:-2.5500000000000003em;margin-left:-0.04398em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.03148em&quot;&gt;k&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.15em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 개별적으로 적용된다.&lt;/li&gt;&lt;li&gt;히든 벡터 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;H&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;H&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.08125em&quot;&gt;H&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 다시 가중치 행렬 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^{(2)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8879999999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;와 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;b^{(2)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8879999999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 통해 선형변환해서 출력한다.&lt;/li&gt;&lt;/ol&gt;&lt;h1 id=&quot;-21&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-21&quot; aria-label=&quot; 21 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:35.546875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA9klEQVQY021Q2W7EIAzM//9hN+peVVKFdAMJhw3m6KTpqqq0foDBZjxjd601yYU4ccrAgVOttf3EcZnNaedTSswxppRzAY4x4luHci4VnBAl5xqliMg4jpwiEU3TtJjtMqrNWmNWpeZh+FwW/XgspZTuUNaWjOPA4mnverlcnXNvp9M0qWEYmUgphXwpFcB7P89fv8pJsqfILLttilBGjZm11tZarQ1E1nXF2f7HQS6Q1ZZtSJ4FCn3fhxCut9v7+UzEkPpbxBPsZDwcRevBhGbZXQTGeNjK/f6xbRjWwUh7FR28oQZvR8v6XHISQQa7xRQvmSL5G4FFlcniVxWyAAAAAElFTkSuQmCC&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;multi-layer-nn&quot; title=&quot;multi-layer-nn&quot; src=&quot;/static/e75ddc1bce937c8b33a3e94a220c5ee0/2bef9/multi-layer-nn.png&quot; srcSet=&quot;/static/e75ddc1bce937c8b33a3e94a220c5ee0/6f3f2/multi-layer-nn.png 256w,/static/e75ddc1bce937c8b33a3e94a220c5ee0/01e7c/multi-layer-nn.png 512w,/static/e75ddc1bce937c8b33a3e94a220c5ee0/2bef9/multi-layer-nn.png 1024w,/static/e75ddc1bce937c8b33a3e94a220c5ee0/71c1d/multi-layer-nn.png 1536w,/static/e75ddc1bce937c8b33a3e94a220c5ee0/a878e/multi-layer-nn.png 2048w,/static/e75ddc1bce937c8b33a3e94a220c5ee0/09096/multi-layer-nn.png 2104w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-22&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-22&quot; aria-label=&quot; 22 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;다층(multi-layer) 퍼셉트론, MLP&lt;/code&gt;&lt;/strong&gt;는 이처럼 &lt;strong&gt;[선형변환 - 활성함수 합성] 사이클을 n회 반복하여 신경망이 여러층 합성된 함수&lt;/strong&gt;이다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;MLP의 파라미터는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;L&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;개의 가중치 행렬 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^{(L)},\dots,W^{(1)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0824399999999998em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;과 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;절편 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;b^{(L)},\dots,b^{(1)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.0824399999999998em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;L&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;로 이루어져 있다.&lt;/li&gt;&lt;li&gt;이 때 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;ℓ&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mo&gt;…&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mi&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\ell = 1,\dots,L&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;ℓ&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8777699999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;…&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt; 까지의 순차적인 신경망 계산을 &lt;strong&gt;&lt;code&gt;순전파(forward propagation)&lt;/code&gt;&lt;/strong&gt;이라고 부른다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-23&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-23&quot; aria-label=&quot; 23 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;층을-여러개-쌓는-이유&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%B8%B5%EC%9D%84-%EC%97%AC%EB%9F%AC%EA%B0%9C-%EC%8C%93%EB%8A%94-%EC%9D%B4%EC%9C%A0&quot; aria-label=&quot;층을 여러개 쌓는 이유 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;층을 여러개 쌓는 이유&lt;/h3&gt;&lt;p&gt;이론적으로는 2층 신경망으로도 임의의 연속함수를 근사할 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이를 &lt;code&gt;universal approximation theorem&lt;/code&gt;이라고 한다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-24&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-24&quot; aria-label=&quot; 24 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;그러나 층이 깊을수록 &lt;strong&gt;목적함수를 근사하는데 필요한 뉴런(노드)의 숫자가 훨씬 빨리 줄어들어 좀 더 효율적으로 학습이 가능&lt;/strong&gt;하다. 즉, 바꾸어 말하면 &lt;strong&gt;&lt;div&gt;적은 수의 뉴런으로도 층을 더 깊게 쌓으면 훨씬 더 복잡한 함수를 표현할 수 있다.&lt;/div&gt;&lt;/strong&gt;&lt;/p&gt;&lt;ul&gt;&lt;li&gt;층이 얇으면 필요한 뉴런의 숫자가 기하급수적으로 늘어나, 넓은(wide) 신경망이 되어야한다.&lt;/li&gt;&lt;/ul&gt;&lt;div&gt;&lt;p&gt;주의할 점은, 층이 깊을수록 최적화가 더 어려워져 학습하기가 어려워진다는 것이다. 이 부분에 대해서는 합성곱신경망(Convolution Neural Network) 파트의 Residual Block에서 자세히 설명한다.&lt;/p&gt;&lt;/div&gt;&lt;h1 id=&quot;-25&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-25&quot; aria-label=&quot; 25 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;딥러닝의-학습-원리&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%94%A5%EB%9F%AC%EB%8B%9D%EC%9D%98-%ED%95%99%EC%8A%B5-%EC%9B%90%EB%A6%AC&quot; aria-label=&quot;딥러닝의 학습 원리 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;딥러닝의 학습 원리&lt;/h2&gt;&lt;h3 id=&quot;역전파-알고리즘&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%97%AD%EC%A0%84%ED%8C%8C-%EC%95%8C%EA%B3%A0%EB%A6%AC%EC%A6%98&quot; aria-label=&quot;역전파 알고리즘 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;역전파 알고리즘&lt;/h3&gt;&lt;p&gt;앞서 말했던 &lt;strong&gt;&lt;code&gt;순전파(forward propagation)&lt;/code&gt;&lt;/strong&gt;는 입력값 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;X&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;X&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07847em&quot;&gt;X&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 받아서 선형모델과 활성함수를 반복적으로 적용하여 출력하는 연산이었다.&lt;/p&gt;&lt;p&gt;이 때, 가중치 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.68333em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 학습시키려면 각각의 가중치에 대한 gradient 벡터를 계산해야한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이는 &lt;a href=&quot;https://blogik.netlify.app/BoostCamp/U_stage/gradient_descent/#%EA%B2%BD%EC%82%AC%ED%95%98%EA%B0%95%EB%B2%95%EC%9C%BC%EB%A1%9C-%EC%84%A0%ED%98%95%ED%9A%8C%EA%B7%80-%EA%B3%84%EC%88%98-%EA%B5%AC%ED%95%98%EA%B8%B0&quot;&gt;선형회귀에서의 경사하강법&lt;/a&gt;에서 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;β&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\beta&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8888799999999999em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.05278em&quot;&gt;β&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 구하던 것과 같은 개념이다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-26&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-26&quot; aria-label=&quot; 26 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이 과정을 &lt;strong&gt;&lt;code&gt;역전파(backpropagation)&lt;/code&gt;&lt;/strong&gt; 알고리즘으로 수행한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;각 층에 존재하는 파라미터들에 대한 미분을 계산해서, 그 미분 값을 가지고 파라미터를 업데이트한다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-27&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-27&quot; aria-label=&quot; 27 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;가중치를 업데이트할 때, &lt;strong&gt;행렬의 모든 원소개수만큼, 또 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;절편의 모든 원소 개수만큼 경사하강법이 적용&lt;/strong&gt;된다. 따라서 기존의 선형모델보다 훨씬 더 많은 파라미터들에 대해 경사하강법을 적용하게 된다.&lt;/p&gt;&lt;p&gt;또, 선형모델은 한 층에 대해서만 계산하는 모델이므로 gradient 벡터를 한번에 계산할 수 있었지만, 딥러닝은 여러 층에 걸쳐 순차적으로 계산하기 때문에 gradient 벡터를 한번에 계산할 수 없다.&lt;/p&gt;&lt;p&gt;따라서 &lt;strong&gt;&lt;code&gt;역전파&lt;/code&gt;&lt;/strong&gt; 알고리즘은 &lt;strong&gt;&lt;code&gt;순전파&lt;/code&gt;&lt;/strong&gt;와 비슷하게 &lt;strong&gt;&lt;div&gt;역순차적으로 층마다 미분을 계산하여 적용&lt;/div&gt;&lt;/strong&gt;시킨다.&lt;/p&gt;&lt;h1 id=&quot;-28&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-28&quot; aria-label=&quot; 28 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;역전파-알고리즘-원리-이해하기&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%97%AD%EC%A0%84%ED%8C%8C-%EC%95%8C%EA%B3%A0%EB%A6%AC%EC%A6%98-%EC%9B%90%EB%A6%AC-%EC%9D%B4%ED%95%B4%ED%95%98%EA%B8%B0&quot; aria-label=&quot;역전파 알고리즘 원리 이해하기 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;역전파 알고리즘 원리 이해하기&lt;/h3&gt;&lt;h1 id=&quot;-29&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-29&quot; aria-label=&quot; 29 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:33.203125%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABQAAAAHCAIAAACHqfpvAAAACXBIWXMAABYlAAAWJQFJUiTwAAAA6UlEQVQY03VQi2rEMAzr/3/irb2tT5akvT5IYzvP6VYGg90EFoljOUJVLiWllHNGlVKIyFq7bhsTlRjRYfbzdoQU/YUQUGBIKhRJkBBDiDGlruuV1vO8KKVl3UpOaDeDmpTph2GeH3haloc2syOCuJzkLXn2kUWa5k7EmBvGUX0qa4zdd2IuPwZ/o0LDOlgLEMOM1sY5N00TeBxHsDEGn5RXqOD1OOWwQhIdcV3XzLyu60fb8ukyuVz+RRVCYBZYwgVn+BWR2+3tfn/f96Nvu/gd2188bSMYBADBdb+2iHhsfLL4l2JMIuAv+0KVbapS40gAAAAASUVORK5CYII=&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;backpropagation_desc&quot; title=&quot;backpropagation_desc&quot; src=&quot;/static/b0fed3550d05be76839fd0e83928bd8f/2bef9/backpropagation_desc.png&quot; srcSet=&quot;/static/b0fed3550d05be76839fd0e83928bd8f/6f3f2/backpropagation_desc.png 256w,/static/b0fed3550d05be76839fd0e83928bd8f/01e7c/backpropagation_desc.png 512w,/static/b0fed3550d05be76839fd0e83928bd8f/2bef9/backpropagation_desc.png 1024w,/static/b0fed3550d05be76839fd0e83928bd8f/71c1d/backpropagation_desc.png 1536w,/static/b0fed3550d05be76839fd0e83928bd8f/a878e/backpropagation_desc.png 2048w,/static/b0fed3550d05be76839fd0e83928bd8f/b3608/backpropagation_desc.png 2160w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;h1 id=&quot;-30&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-30&quot; aria-label=&quot; 30 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;code&gt;손실함수&lt;/code&gt;를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\mathscr{L}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.7em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathscr&quot; style=&quot;margin-right:0.19189em&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이라고 했을 때, 각각의 가중치 행렬 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^{(\mathscr{L})}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8879999999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mathscr mtight&quot; style=&quot;margin-right:0.19189em&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;마다 손실함수에 대한 미분을 계산할 때 &lt;strong&gt;&lt;code&gt;역전파&lt;/code&gt;&lt;/strong&gt; 알고리즘을 사용한다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;이 때, 각 층에서 계산된 gradient 벡터들은 &lt;strong&gt;밑의 층으로 전달&lt;/strong&gt;되는 flow 형태이다.&lt;ul&gt;&lt;li&gt;저층에 있는 gradient 벡터를 계산할 때는 위층에 있는 gradient 벡터가 필요하다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;즉, &lt;strong&gt;위층에서 아래층으로 내려오면서&lt;/strong&gt; 업데이트하는 방식이다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-31&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-31&quot; aria-label=&quot; 31 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;ℓ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;Z&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;ℓ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;ℓ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{\partial\mathcal{L}}{\partial W^{(\ell)}} = \frac{\partial\mathcal{L}}{\partial O}\times\cdots\times\frac{\partial Z^{(\ell)}}{\partial W^{(\ell)}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.07544em;vertical-align:-0.704em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.2960000000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.814em&quot;&gt;&lt;span style=&quot;top:-2.989em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;ℓ&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.704em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.05744em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.66666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.269em;vertical-align:-0.704em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.565em&quot;&gt;&lt;span style=&quot;top:-2.2960000000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.814em&quot;&gt;&lt;span style=&quot;top:-2.989em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;ℓ&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;Z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;ℓ&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.704em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;ℓ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;O&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mo&gt;⋯&lt;/mo&gt;&lt;mo&gt;×&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;Z&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;ℓ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msup&gt;&lt;mi&gt;b&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;ℓ&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{\partial\mathcal{L}}{\partial b^{(\ell)}} = \frac{\partial\mathcal{L}}{\partial O}\times\cdots\times\frac{\partial Z^{(\ell)}}{\partial b^{(\ell)}}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.07544em;vertical-align:-0.704em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.2960000000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.814em&quot;&gt;&lt;span style=&quot;top:-2.989em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;ℓ&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.704em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.05744em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02778em&quot;&gt;O&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.66666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;⋯&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;×&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.269em;vertical-align:-0.704em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.565em&quot;&gt;&lt;span style=&quot;top:-2.2960000000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;b&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.814em&quot;&gt;&lt;span style=&quot;top:-2.989em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;ℓ&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.07153em&quot;&gt;Z&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;ℓ&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.704em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;h1 id=&quot;-32&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-32&quot; aria-label=&quot; 32 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;이 원리를 사용할 때 &lt;strong&gt;합성함수의 미분법&lt;/strong&gt;인 &lt;strong&gt;&lt;code&gt;연쇄법칙&lt;/code&gt;&lt;/strong&gt;을 사용하여 gradient 벡터를 전달한다.&lt;/p&gt;&lt;h1 id=&quot;-33&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-33&quot; aria-label=&quot; 33 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;strong&gt;&lt;code&gt;연쇄법칙(chain-rule)&lt;/code&gt;&lt;/strong&gt;은 합성함수를 미분하는 방식이다. 오늘날의 딥러닝 프레임워크들은 이러한 연쇄법칙을 기반으로 한 &lt;strong&gt;&lt;code&gt;자동 미분(auto-differentiation)&lt;/code&gt;&lt;/strong&gt;을 수행하여 신경망을 학습시킨다.&lt;/p&gt;&lt;h1 id=&quot;-34&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-34&quot; aria-label=&quot; 34 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;연쇄법칙-예제&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EC%97%B0%EC%87%84%EB%B2%95%EC%B9%99-%EC%98%88%EC%A0%9C&quot; aria-label=&quot;연쇄법칙 예제 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;연쇄법칙 예제&lt;/h3&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;msup&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z = (x+y)^2&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.1141079999999999em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;위와 같은 함수가 있을 때, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해서 어떻게 미분할 수 있는가?&lt;/p&gt;&lt;p&gt;일단, 위의 함수를 다음과 같이 두 함수의 결합으로 표현할 수 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;msup&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;/msup&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z = w^2 \\
w = x+y&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8641079999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8641079999999999em&quot;&gt;&lt;span style=&quot;top:-3.113em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;2&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.66666em;vertical-align:-0.08333em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.625em;vertical-align:-0.19444em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이 함수들을 토대로 다음과 같이 연쇄적인 미분값의 곱으로 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{\partial z}{\partial x}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1.2251079999999999em;vertical-align:-0.345em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8801079999999999em&quot;&gt;&lt;span style=&quot;top:-2.6550000000000002em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.394em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal mtight&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.345em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;를 표현할 수 있다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{\partial z}{\partial x} = \frac{\partial z}{\partial w}\frac{\partial w}{\partial x}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.05744em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.05744em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;위의 식의 두 미분값을 따로 떼내어 두 함수에 적용시키면 다음과 같다.&lt;/p&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mspace linebreak=&quot;newline&quot;&gt;&lt;/mspace&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mrow&gt;&lt;mo fence=&quot;true&quot;&gt;(&lt;/mo&gt;&lt;mo separator=&quot;true&quot;&gt;,&lt;/mo&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo fence=&quot;true&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{\partial z}{\partial w}= 2w\\
\frac{\partial w}{\partial x} = 1 \left ( ,\frac{\partial w}{\partial y} = 1 \right )&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.05744em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace newline&quot;&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.05744em;vertical-align:-0.686em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.686em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.40003em;vertical-align:-0.95003em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;minner&quot;&gt;&lt;span class=&quot;mopen delimcenter&quot; style=&quot;top:0em&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;(&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mpunct&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.16666666666666666em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.314em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8804400000000001em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose delimcenter&quot; style=&quot;top:0em&quot;&gt;&lt;span class=&quot;delimsizing size3&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;이 두 미분값의 곱, &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mi&gt;w&lt;/mi&gt;&lt;mo&gt;⋅&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mn&gt;2&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;mo&gt;+&lt;/mo&gt;&lt;mi&gt;y&lt;/mi&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;2w \cdot 1 = 2(x+y)&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.02691em&quot;&gt;w&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;⋅&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.64444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;mopen&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mbin&quot;&gt;+&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2222222222222222em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:1em;vertical-align:-0.25em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.03588em&quot;&gt;y&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;이  &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;z&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;z&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.04398em&quot;&gt;z&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;의 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mi&gt;x&lt;/mi&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;x&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.43056em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord mathnormal&quot;&gt;x&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대한 미분값이다.&lt;/p&gt;&lt;p&gt;위와 같은 합성함수미분의 연쇄법칙 과정을 &lt;strong&gt;&lt;code&gt;역전파&lt;/code&gt;&lt;/strong&gt;에 적용시킬 수 있다.&lt;/p&gt;&lt;div&gt;&lt;p&gt;단, 이 방식은 각 뉴런의 값(텐서 값)이 컴퓨터 메모리에 저장되어야 역전파 알고리즘에 사용할 수 있다. 즉, x에 대해 미분하고싶다면 x와 y값을 알고있어야만 미분 가능하다. 따라서 역전파는 순전파보다 다소 메모리를 많이 사용하게 되는 방법이다.&lt;/p&gt;&lt;/div&gt;&lt;h1 id=&quot;-35&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-35&quot; aria-label=&quot; 35 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;2층-신경망-예제&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#2%EC%B8%B5-%EC%8B%A0%EA%B2%BD%EB%A7%9D-%EC%98%88%EC%A0%9C&quot; aria-label=&quot;2층 신경망 예제 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;2층 신경망 예제&lt;/h3&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:46.09375%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;2-layer-nn-ex&quot; title=&quot;2-layer-nn-ex&quot; src=&quot;/static/c3146bdfa66955df72ee2114b747a348/2bef9/2-layer-nn-ex.png&quot; srcSet=&quot;/static/c3146bdfa66955df72ee2114b747a348/6f3f2/2-layer-nn-ex.png 256w,/static/c3146bdfa66955df72ee2114b747a348/01e7c/2-layer-nn-ex.png 512w,/static/c3146bdfa66955df72ee2114b747a348/2bef9/2-layer-nn-ex.png 1024w,/static/c3146bdfa66955df72ee2114b747a348/71c1d/2-layer-nn-ex.png 1536w,/static/c3146bdfa66955df72ee2114b747a348/a878e/2-layer-nn-ex.png 2048w,/static/c3146bdfa66955df72ee2114b747a348/1ffbd/2-layer-nn-ex.png 2108w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;위에서 배운 연쇄법칙을 이용하여 2층 신경망에서의 역전파 과정을 수식으로 정리해볼 수 있다.&lt;/p&gt;&lt;p&gt;위의 그림에서, &lt;strong&gt;파란 화살표&lt;/strong&gt;는 &lt;strong&gt;&lt;code&gt;순전파&lt;/code&gt;&lt;/strong&gt;, &lt;strong&gt;빨간 화살표&lt;/strong&gt;는 &lt;strong&gt;&lt;code&gt;역전파&lt;/code&gt;&lt;/strong&gt;를 의미한다.  &lt;/p&gt;&lt;h1 id=&quot;-36&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-36&quot; aria-label=&quot; 36 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;span class=&quot;katex-display&quot;&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot; display=&quot;block&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;mfrac&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;mi mathvariant=&quot;script&quot;&gt;L&lt;/mi&gt;&lt;/mrow&gt;&lt;mrow&gt;&lt;mi mathvariant=&quot;normal&quot;&gt;∂&lt;/mi&gt;&lt;msup&gt;&lt;mi mathvariant=&quot;bold&quot;&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;/mfrac&gt;&lt;mo&gt;=&lt;/mo&gt;&lt;mo stretchy=&quot;false&quot;&gt;?&lt;/mo&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;\frac{\partial\mathcal{L}}{\partial\bold{W}^{(1)}} = ?&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:2.07544em;vertical-align:-0.704em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mopen nulldelimiter&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mfrac&quot;&gt;&lt;span class=&quot;vlist-t vlist-t2&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:1.37144em&quot;&gt;&lt;span style=&quot;top:-2.2960000000000003em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathbf&quot; style=&quot;margin-right:0.01597em&quot;&gt;W&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.814em&quot;&gt;&lt;span style=&quot;top:-2.989em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.23em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;frac-line&quot; style=&quot;border-bottom-width:0.04em&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span style=&quot;top:-3.677em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:3em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord&quot; style=&quot;margin-right:0.05556em&quot;&gt;∂&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathcal&quot;&gt;L&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-s&quot;&gt;​&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.704em&quot;&gt;&lt;span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mclose nulldelimiter&quot;&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;mspace&quot; style=&quot;margin-right:0.2777777777777778em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mrel&quot;&gt;=&lt;/span&gt;&lt;/span&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.69444em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mclose&quot;&gt;?&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;p&gt;첫번째 층에 해당하는 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^{(1)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8879999999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 대해서 경사하강법을 사용하고 싶을 때, 어떻게 gradient 벡터를 계산하는가?&lt;/p&gt;&lt;ul&gt;&lt;li&gt;&lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^{(1)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8879999999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;은 행렬이므로 각 성분에 대한 편미분을 계산해주어야한다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-37&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-37&quot; aria-label=&quot; 37 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;각각의 연쇄법칙을 순서대로 적용할 때, 미분값이 각 층마다 계산되므로, 경사하강법에서 실제 가중치 행렬 &lt;span class=&quot;katex&quot;&gt;&lt;span class=&quot;katex-mathml&quot;&gt;&lt;math xmlns=&quot;http://www.w3.org/1998/Math/MathML&quot;&gt;&lt;semantics&gt;&lt;mrow&gt;&lt;msup&gt;&lt;mi&gt;W&lt;/mi&gt;&lt;mrow&gt;&lt;mo stretchy=&quot;false&quot;&gt;(&lt;/mo&gt;&lt;mn&gt;1&lt;/mn&gt;&lt;mo stretchy=&quot;false&quot;&gt;)&lt;/mo&gt;&lt;/mrow&gt;&lt;/msup&gt;&lt;/mrow&gt;&lt;annotation encoding=&quot;application/x-tex&quot;&gt;W^{(1)}&lt;/annotation&gt;&lt;/semantics&gt;&lt;/math&gt;&lt;/span&gt;&lt;span class=&quot;katex-html&quot; aria-hidden=&quot;true&quot;&gt;&lt;span class=&quot;base&quot;&gt;&lt;span class=&quot;strut&quot; style=&quot;height:0.8879999999999999em;vertical-align:0em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;mord&quot;&gt;&lt;span class=&quot;mord mathnormal&quot; style=&quot;margin-right:0.13889em&quot;&gt;W&lt;/span&gt;&lt;span class=&quot;msupsub&quot;&gt;&lt;span class=&quot;vlist-t&quot;&gt;&lt;span class=&quot;vlist-r&quot;&gt;&lt;span class=&quot;vlist&quot; style=&quot;height:0.8879999999999999em&quot;&gt;&lt;span style=&quot;top:-3.063em;margin-right:0.05em&quot;&gt;&lt;span class=&quot;pstrut&quot; style=&quot;height:2.7em&quot;&gt;&lt;/span&gt;&lt;span class=&quot;sizing reset-size6 size3 mtight&quot;&gt;&lt;span class=&quot;mord mtight&quot;&gt;&lt;span class=&quot;mopen mtight&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;mord mtight&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;mclose mtight&quot;&gt;)&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;에 적용시킬 gradient 벡터를 찾아낼 수 있다.&lt;/p&gt;&lt;p&gt;딥러닝을 학습시킬때는 이렇게 계산한 각각의 가중치 행렬에 대한 gradient 벡터를  &lt;strong&gt;&lt;code&gt;SGD&lt;/code&gt;&lt;/strong&gt;에 이용하여, &lt;strong&gt;데이터를 바꾸어가면서(mini batch) 파라미터들을 학습&lt;/strong&gt;시킨다. 이로써 주어진 목적식을 최소화하는 파라미터들을 찾을 수 있다.&lt;/p&gt;&lt;h1 id=&quot;-38&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-38&quot; aria-label=&quot; 38 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;레퍼런스&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#%EB%A0%88%ED%8D%BC%EB%9F%B0%EC%8A%A4&quot; aria-label=&quot;레퍼런스 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;레퍼런스&lt;/h3&gt;&lt;p&gt;순전파와 역전파의 개념을 좀 더 명확히 알고싶다면, 다음 레퍼런스를 참고하자.&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://d2l.ai/chapter_multilayer-perceptrons/backprop.html&quot;&gt;4.7. Forward Propagation, Backward Propagation, and Computational Graphs - Dive into Deep Learning 0.16.1 documentation&lt;/a&gt;&lt;/p&gt;&lt;p&gt;&lt;a href=&quot;https://ko.d2l.ai/chapter_deep-learning-basics/backprop.html&quot;&gt;3.14. 순전파(forward propagation), 역전파(back propagation), 연산 그래프 - Dive into Deep Learning documentation&lt;/a&gt;&lt;/p&gt;</content:encoded></item><item><title><![CDATA[Pandas 라이브러리 사용법]]></title><description><![CDATA[Pandas for AI Math by 최성철 교수님, BoostCamp AI Tech 2주차]]></description><link>https://blogik.netlify.app/BoostCamp/U_stage/06_pandas/</link><guid isPermaLink="false">https://blogik.netlify.app/BoostCamp/U_stage/06_pandas/</guid><pubDate>Wed, 27 Jan 2021 00:00:00 GMT</pubDate><content:encoded>&lt;p&gt;본 정리 내용은 &lt;a href=&quot;https://boostcamp.connect.or.kr/&quot;&gt;Naver BoostCamp AI Tech&lt;/a&gt;의 edwith에서 수강한 내용을 정리한 것입니다.&lt;br/&gt;
사실과 다른 부분이 있거나, 수정이 필요한 사항은 댓글로 남겨주세요.&lt;/p&gt;&lt;hr/&gt;&lt;h1 id=&quot;pandas&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#pandas&quot; aria-label=&quot;pandas permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Pandas&lt;/h1&gt;&lt;p&gt;&lt;code&gt;panel datas&lt;/code&gt;에서 나온 이름으로, &lt;strong&gt;구조화된 데이터의 처리를 지원&lt;/strong&gt;하는 Python의 라이브러리이다. Python의 엑셀이라고 할 수 있다.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;고성능 array 계산용 라이브러리인 numpy와 통합하여 강력한 스프레드시트 처리 기능을 제공한다.&lt;/li&gt;&lt;li&gt;인덱싱, 연산용 함수, 전처리 함수 등을 제공한다.&lt;/li&gt;&lt;li&gt;데이터 처리 및 통계 분석을 위해 사용된다.&lt;/li&gt;&lt;li&gt;Tabular 데이터(테이블형 데이터)를 다루는 데에 최적화되어있다.&lt;/li&gt;&lt;/ul&gt;&lt;p&gt;&lt;span class=&quot;gatsby-resp-image-wrapper&quot; style=&quot;position:relative;display:block;margin-left:auto;margin-right:auto;max-width:1024px&quot;&gt;
      &lt;span class=&quot;gatsby-resp-image-background-image&quot; style=&quot;padding-bottom:54.296875%;position:relative;bottom:0;left:0;background-image:url(&amp;#x27;data:image/png;base64,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&amp;#x27;);background-size:cover;display:block&quot;&gt;&lt;/span&gt;
  &lt;img class=&quot;gatsby-resp-image-image&quot; alt=&quot;pandas_data_type&quot; title=&quot;pandas_data_type&quot; src=&quot;/static/1d53b1562735aed61e79f68bd0de2eb4/2bef9/pandas_data_type.png&quot; srcSet=&quot;/static/1d53b1562735aed61e79f68bd0de2eb4/6f3f2/pandas_data_type.png 256w,/static/1d53b1562735aed61e79f68bd0de2eb4/01e7c/pandas_data_type.png 512w,/static/1d53b1562735aed61e79f68bd0de2eb4/2bef9/pandas_data_type.png 1024w,/static/1d53b1562735aed61e79f68bd0de2eb4/71c1d/pandas_data_type.png 1536w,/static/1d53b1562735aed61e79f68bd0de2eb4/a878e/pandas_data_type.png 2048w,/static/1d53b1562735aed61e79f68bd0de2eb4/7f486/pandas_data_type.png 3010w&quot; sizes=&quot;(max-width: 1024px) 100vw, 1024px&quot; style=&quot;width:100%;height:100%;margin:0;vertical-align:middle;position:absolute;top:0;left:0&quot; loading=&quot;lazy&quot;/&gt;
    &lt;/span&gt;&lt;/p&gt;&lt;p&gt;이 중, Data Table 전체를 포함하는 객체를 &lt;strong&gt;&lt;code&gt;DataFrame&lt;/code&gt;&lt;/strong&gt;, Column 하나에 해당하는 데이터셋 객체를 &lt;strong&gt;&lt;code&gt;Series&lt;/code&gt;&lt;/strong&gt;라고 한다.&lt;/p&gt;&lt;h2 id=&quot;series&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#series&quot; aria-label=&quot;series permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Series&lt;/h2&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;python&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-python&quot;&gt;PYTHON&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-python&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;list_data &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;3&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;4&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;example_obj &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; Series&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;data &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; list_data&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;example_obj&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;&amp;#x27;&amp;#x27;&amp;#x27;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;0    1&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;1    2&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;2    3&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;3    4&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;4    5&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;dtype: int64&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;&amp;#x27;&amp;#x27;&amp;#x27;&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;h1 id=&quot;&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#&quot; aria-label=&quot; permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;p&gt;&lt;code&gt;data&lt;/code&gt; 파라미터를 받아 객체를 생성하며, &lt;code&gt;Numpy&lt;/code&gt; 의 &lt;code&gt;ndarray&lt;/code&gt; 를 기반으로 만들어진, &lt;code&gt;Pandas&lt;/code&gt; 에서 사용하기 위한 wrapper 객체라고 볼수도 있다.(Subclass of &lt;code&gt;numpy.ndarray&lt;/code&gt;)&lt;/p&gt;&lt;p&gt;특이한 점은, 위와 같이 &lt;code&gt;[1,2,3,4,5]&lt;/code&gt; 라는 데이터만 들어가는것이 아니라, 이 데이터의 인덱스 값 &lt;code&gt;[0,1,2,3,4]&lt;/code&gt; 도 같이 저장된다. 이 인덱스값은 &lt;strong&gt;&lt;div&gt;문자열로도 지정 가능&lt;/div&gt;&lt;/strong&gt;하다.&lt;/p&gt;&lt;div&gt;&lt;p&gt;Series 객체를 생성할 때, Index를 기준으로 생성하므로 만약 data가 없는 index가 있다면 NaN값이 채워진다.&lt;/p&gt;&lt;/div&gt;&lt;ul&gt;&lt;li&gt;data와 index를 따로 명시하는 대신 dict 타입을 넣어 한번에 지정할 수도 있다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;name&lt;/code&gt;으로 &lt;code&gt;Series&lt;/code&gt; 객체 자체의 이름을 정해줄 수 있다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;index.name&lt;/code&gt; 으로 인덱스의 이름을 지정해줄 수 있다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;as_type&lt;/code&gt; 으로 데이터 타입을 변경할 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-1&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-1&quot; aria-label=&quot; 1 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;python&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-python&quot;&gt;PYTHON&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-python&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token comment&quot;&gt;# 인덱스 문자열로 바꾸기&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;list_data &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;3&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;4&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;list_name &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;a&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;b&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;c&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;d&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;e&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;example_obj &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; Series&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;data &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; list_data&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; index&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;list_name&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;example_obj&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;&amp;#x27;&amp;#x27;&amp;#x27;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;a    1&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;b    2&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;c    3&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;d    4&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;e    5&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;dtype: int64&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;&amp;#x27;&amp;#x27;&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# dict 타입으로 생성하기&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;dict_data &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;a&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;b&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;c&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;3&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;d&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;4&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;e&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;5&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;example_obj &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; Series&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;dict_data&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; dtype&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;np&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;float32&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; name&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;example_data&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# series 객체 이름 지정&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;example_obj&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;name &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;number&amp;quot;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# index 이름 지정&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;example_obj&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;index&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;name &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;alphabet&amp;quot;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 데이터 타입(dtype) 변경하기&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;example_obj &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; example_obj&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;astype&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token builtin&quot;&gt;int&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;h1 id=&quot;-2&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-2&quot; aria-label=&quot; 2 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;dataframe&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#dataframe&quot; aria-label=&quot;dataframe permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Dataframe&lt;/h2&gt;&lt;p&gt;&lt;code&gt;Numpy&lt;/code&gt; 의 2차원 &lt;code&gt;ndarray&lt;/code&gt; 와 비슷한 객체이며, &lt;code&gt;Series&lt;/code&gt; 와 달리 (row) index 뿐만 아니라 &lt;strong&gt;column index&lt;/strong&gt;도 있다. 각 컬럼은 &lt;strong&gt;다른 데이터 타입&lt;/strong&gt;을 가질 수 있으며, 컬럼을 떼고 붙일 수 있으므로 &lt;strong&gt;데이터프레임의 전체 사이즈는 mutable&lt;/strong&gt;하다.&lt;/p&gt;&lt;div&gt;&lt;p&gt;Series 객체를 생성할 때처럼, Index를 기준으로 생성하므로 만약 data가 없는 row 또는 column index가 있다면 NaN값이 채워진다.&lt;/p&gt;&lt;/div&gt;&lt;ul&gt;&lt;li&gt;인덱싱으로 &lt;code&gt;loc&lt;/code&gt; 과 &lt;code&gt;iloc&lt;/code&gt; 을 사용할 수 있다. (&lt;code&gt;Series&lt;/code&gt; 객체도 통용)&lt;ul&gt;&lt;li&gt;&lt;code&gt;loc&lt;/code&gt; 은 실제 index name과 일치하게 사용해야한다. (숫자, 문자열)&lt;/li&gt;&lt;li&gt;&lt;code&gt;iloc&lt;/code&gt; 은 index의 position 값을 사용한다.(0부터 n까지의 숫자)&lt;/li&gt;&lt;li&gt;두 인덱싱 모두 row와 column을 모두 지정해줄 수 있다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;column에 새로운 데이터를 할당할 수 있다.&lt;ul&gt;&lt;li&gt;comparison 등으로 boolean 값을 넣을 수도 있다. (&lt;code&gt;Numpy&lt;/code&gt; 의 &lt;code&gt;fancy index&lt;/code&gt; )&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;df.T&lt;/code&gt; 로 row와 column 인덱스를 바꿀 수 있다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;values&lt;/code&gt; 를 사용하여 데이터 &lt;code&gt;ndarray&lt;/code&gt;를 뽑아낼 수 있다.&lt;/li&gt;&lt;li&gt;컬럼을 삭제할 때, 두가지 방법을 이용할 수있다.&lt;ul&gt;&lt;li&gt;컬럼 삭제 시, &lt;code&gt;del df[&amp;#x27;컬럼명&amp;#x27;]&lt;/code&gt; - 메모리 주소까지 삭제&lt;/li&gt;&lt;li&gt;&lt;code&gt;df.drop(&amp;#x27;컬럼명&amp;#x27;, axis=n)&lt;/code&gt; 원본 데이터 프레임은 변경하지 않고, 해당 컬럼을 삭제한 데이터프레임을 반환한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-3&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-3&quot; aria-label=&quot; 3 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;python&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-python&quot;&gt;PYTHON&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-python&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;raw_data &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;{&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;first_name&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Jason&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Molly&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Tina&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Jake&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Amy&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;        &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;last_name&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Miller&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Jacobson&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Ali&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Milner&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Cooze&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;        &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;age&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;42&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;52&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;36&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;24&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;73&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;        &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;city&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;San Francisco&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Baltimore&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Miami&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Douglas&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;Boston&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;}&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;df &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; pd&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;DataFrame&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;raw_data&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; columns &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;first_name&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;last_name&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;age&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;city&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;df&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;first_name &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# == df[&amp;quot;first_name&amp;quot;], series 데이터형&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 인덱싱&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;s &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; pd&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;Series&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;np&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;nan&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; index&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;49&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;48&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;loc&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 인덱스 이름이 1인 row 까지만&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;&amp;#x27;&amp;#x27;&amp;#x27;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;49   NaN&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;48   NaN&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;1    NaN&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;&amp;#x27;&amp;#x27;&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;s&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;iloc&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;1&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 인덱스 위치가 1인 row까지만&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;&amp;#x27;&amp;#x27;&amp;#x27;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;49   NaN&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token triple-quoted-string string&quot;&gt;&amp;#x27;&amp;#x27;&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# column에 새로운 데이터 할당&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;df&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;debt &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; df&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;age &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;&amp;gt;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;40&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# column &amp;#x27;debt&amp;#x27;에 comparison의 T/F 값 입력하여 붙임&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;h1 id=&quot;-4&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-4&quot; aria-label=&quot; 4 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h2 id=&quot;selection--drop&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#selection--drop&quot; aria-label=&quot;selection  drop permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Selection &amp;amp; Drop&lt;/h2&gt;&lt;h3 id=&quot;selection-방식&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#selection-%EB%B0%A9%EC%8B%9D&quot; aria-label=&quot;selection 방식 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Selection 방식&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;데이터의 타입을 체킹할 때 &lt;code&gt;head().T&lt;/code&gt; 를 사용하여 컬럼에 올바른 값이 들어갔는지 확인하기도 한다.&lt;/li&gt;&lt;li&gt;n개의 컬럼명을 리스트로 지정하여 해당 컬럼 데이터들을 뽑아올 수 있다.&lt;/li&gt;&lt;li&gt;컬럼명을 지정하지 않았을 시, index number로 해당 row를 뽑아올 수 있다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;Numpy&lt;/code&gt;의 &lt;code&gt;fancy index&lt;/code&gt;처럼, 조건식을 넣어 사용할 수도 있다.&lt;/li&gt;&lt;li&gt;Index name 변경도 가능하다.&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-5&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-5&quot; aria-label=&quot; 5 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;python&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-python&quot;&gt;PYTHON&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-python&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token comment&quot;&gt;# 컬럼 데이터 타입 확인&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;df&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;head&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;3&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;T&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 컬럼명 지정&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;df&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;a&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;head&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;df&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;a&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;b&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;c&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;head&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;3&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 컬럼명 없이 사용하는 인덱스 지정은 row를 표시한다.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;df&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;3&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# == df[[0,1,2]], 0,1,2 row 표시&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 조건식 넣어 selection&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;df&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;20&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;&amp;lt;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;age&amp;quot;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;&amp;lt;&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;40&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;h1 id=&quot;-6&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-6&quot; aria-label=&quot; 6 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;loc-iloc-을-활용한-selection&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#loc-iloc-%EC%9D%84-%ED%99%9C%EC%9A%A9%ED%95%9C-selection&quot; aria-label=&quot;loc iloc 을 활용한 selection permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;code&gt;loc&lt;/code&gt;, &lt;code&gt;iloc&lt;/code&gt; 을 활용한 selection&lt;/h3&gt;&lt;p&gt;&lt;code&gt;loc&lt;/code&gt;, &lt;code&gt;iloc&lt;/code&gt;을 활용한 selection도 확인해보자.&lt;/p&gt;&lt;ul&gt;&lt;li&gt;기본 방식&lt;ul&gt;&lt;li&gt;&lt;strong&gt;컬럼명&lt;/strong&gt;을 명시하고, &lt;strong&gt;index number&lt;/strong&gt;를 넣어 row의 개수를 지정한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;loc&lt;/code&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;index name&lt;/strong&gt;을 넣어 row를 지정한 뒤, &lt;strong&gt;컬럼명&lt;/strong&gt;을 명시한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;li&gt;&lt;code&gt;iloc&lt;/code&gt;&lt;ul&gt;&lt;li&gt;&lt;strong&gt;컬럼 number&lt;/strong&gt;를 지정한 뒤, &lt;strong&gt;index number&lt;/strong&gt;를 지정해 row도 명시한다.&lt;/li&gt;&lt;/ul&gt;&lt;/li&gt;&lt;/ul&gt;&lt;h1 id=&quot;-7&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-7&quot; aria-label=&quot; 7 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;python&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-python&quot;&gt;PYTHON&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-python&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token comment&quot;&gt;# basic&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 컬럼명과 row index 숫자&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;df&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;name&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;street&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# loc&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# row index 명과 컬럼명&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;df&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;loc&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;c&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;#x27;e&amp;#x27;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;name&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token string&quot;&gt;&amp;quot;street&amp;quot;&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# iloc&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 컬럼 숫자와 row index 숫자&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;df&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;iloc&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;[&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;:&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;2&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;]&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;h1 id=&quot;-8&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-8&quot; aria-label=&quot; 8 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;index-재구성&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#index-%EC%9E%AC%EA%B5%AC%EC%84%B1&quot; aria-label=&quot;index 재구성 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Index 재구성&lt;/h3&gt;&lt;p&gt;두 가지 방식이 있다.&lt;/p&gt;&lt;ol&gt;&lt;li&gt;직접 지정&lt;/li&gt;&lt;li&gt;&lt;code&gt;reset_index&lt;/code&gt; 함수 사용&lt;/li&gt;&lt;/ol&gt;&lt;h1 id=&quot;-9&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-9&quot; aria-label=&quot; 9 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;div class=&quot;gatsby-highlight&quot; data-language=&quot;python&quot;&gt;&lt;div class=&quot;badge-btn-wrap&quot;&gt;&lt;div class=&quot;language-badge language-badge-python&quot;&gt;PYTHON&lt;/div&gt;&lt;button class=&quot;btn-copy&quot;&gt;Copy&lt;/button&gt;&lt;/div&gt;&lt;pre class=&quot;prism-code language-python&quot;&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token comment&quot;&gt;# 직접 지정해주기&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;df&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;index &lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt; &lt;/span&gt;&lt;span class=&quot;token builtin&quot;&gt;list&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token builtin&quot;&gt;range&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;0&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;,&lt;/span&gt;&lt;span class=&quot;token number&quot;&gt;10&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 디폴트 인덱스를 추가한 데이터 프레임을 반환&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 기존 인덱스를 대체하고싶다면 drop=True 옵션&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;df&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;.&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;reset_index&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;(&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;drop&lt;/span&gt;&lt;span class=&quot;token operator&quot;&gt;=&lt;/span&gt;&lt;span class=&quot;token boolean&quot;&gt;True&lt;/span&gt;&lt;span class=&quot;token punctuation&quot;&gt;)&lt;/span&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;/div&gt;&lt;div class=&quot;token-line&quot;&gt;&lt;span class=&quot;token plain&quot;&gt;&lt;/span&gt;&lt;span class=&quot;token comment&quot;&gt;# 단, inplace=True 옵션을 사용하면 원본 데이터 프레임 변환을 &amp;#x27;적용&amp;#x27;시킨다.&lt;/span&gt;&lt;/div&gt;&lt;/pre&gt;&lt;/div&gt;&lt;h1 id=&quot;-10&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-10&quot; aria-label=&quot; 10 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;div&gt;&lt;p&gt;일반적으로 데이터프레임은 원본을 유지하는것이 컨벤션이지만, 부득이하게 변경해야할 경우 inplace=True 옵션을 사용하여 변환상태를 &amp;#x27;적용&amp;#x27;시킨다.&lt;/p&gt;&lt;/div&gt;&lt;h1 id=&quot;-11&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#-11&quot; aria-label=&quot; 11 permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;&lt;/h1&gt;&lt;h3 id=&quot;drop&quot; style=&quot;position:relative&quot;&gt;&lt;a href=&quot;#drop&quot; aria-label=&quot;drop permalink&quot; class=&quot;anchor-heading before&quot;&gt;&lt;svg aria-hidden=&quot;true&quot; focusable=&quot;false&quot; height=&quot;16&quot; version=&quot;1.1&quot; viewBox=&quot;0 0 16 16&quot; width=&quot;16&quot;&gt;&lt;path fill-rule=&quot;evenodd&quot; d=&quot;M4 9h1v1H4c-1.5 0-3-1.69-3-3.5S2.55 3 4 3h4c1.45 0 3 1.69 3 3.5 0 1.41-.91 2.72-2 3.25V8.59c.58-.45 1-1.27 1-2.09C10 5.22 8.98 4 8 4H4c-.98 0-2 1.22-2 2.5S3 9 4 9zm9-3h-1v1h1c1 0 2 1.22 2 2.5S13.98 12 13 12H9c-.98 0-2-1.22-2-2.5 0-.83.42-1.64 1-2.09V6.25c-1.09.53-2 1.84-2 3.25C6 11.31 7.55 13 9 13h4c1.45 0 3-1.69 3-3.5S14.5 6 13 6z&quot;&gt;&lt;/path&gt;&lt;/svg&gt;&lt;/a&gt;Drop&lt;/h3&gt;&lt;ul&gt;&lt;li&gt;&lt;code&gt;df.drop(n)&lt;/code&gt; 은 인덱스를 지정하여 해당 row를 삭제한 데이터프레임을 반환한다.&lt;/li&gt;&lt;li&gt;&lt;code&gt;axis=1&lt;/code&gt; 옵션을 지정하여 column을 삭제할 수 있다.&lt;/li&gt;&lt;li&gt;마찬가지로