Open chullhwan-song opened 6 years ago
우리는 수식 1를 optimze하기 위해서 다음 알고리즘(back-pro)을 수행한다.
실제 GAN의 목적은 D가 G를 위조라고 판별하도록 하는게 아니라, G가 D를 속일수 있도록 학습하는게 최종 목적이다.
이 조건이 바로,
앞에서 잠깐 언급했떤거와 같이 이 조건이 만족하기위한 수식 이고 이는 1/2 값을 가진다.
이를 증명하는 것은 학습동안 주어진 G에 대해 D가 를 maximize하는 것이다.
다시, 수식 1를 위의 개념( =G가 D를 속일수 있도록 학습) 으로 표현하면,
이제 다시 생각해보면, 어떤 데이터 셋 (도메인) X가 있고, 어떤 Z가 어떤 행위(G)를 통해 X와 비슷해지는 행위라 정의된다. 다시, X의 확률 분포와 G(Z)확률 분포가 비슷해지는 개념(?) > X의 분포와 유사한 확률분포를 찾는 What?/개념 > 이는 data(X)와 model(G(Z))사이의 KL divergence를 최소화하는 것을 의미 > 보통 이런 개념에선 KL divergence 개념이.ㅎ
https://github.com/yjucho1/articles/edit/master/fromGANtoWGAN/readme.md 참고하여 설명.
일단 위의 개념을 보고, ㅎ 100% 설명에 자신 없으므로 일단 업데이트 할때까지 논문을 발췌, 바로 위의 개념을 다음과 같이 표현가능하다.
G와 D 모두 최적값일때, pg = pr이고, D*(x)=1/2가 되어 손실함수는 -2log2
JS divergence를 적용하여 pr과 pg 사이의 JS divergence는 아래처럼 계산
위식을 다음과 같이 풀어본다.