Adversarial Auto-Encoders

  • Distribution of hidden code

48-adversarialautoencoders-encoder.jpg

  • Give more details after GAN

48-adversarialautoencoders-gan.jpg

Another Approach: q(z)->p(z)

  • Explicitly enforce

48-adversarialautoencoders-kl.jpg

Intuitively comprehend KL(p|q)

\(KL \in{[0,+\infty]}\)

  • KL = 0 --> \(p\approx{q}\)

48-adversarialautoencoders-klimg.jpg

Minimize KL Divergence

  • Evidence Lower BOund

\[KL(q_\theta(z|x_i)||p(z)) \]

48-adversarialautoencoders-minkl.jpg

How to compute KL between q(z) and p(z)

48-adversarialautoencoders-computerkl.jpg

我们所希望的两点:

  • 重建变量
  • 隐藏变量逼近p的分布
posted @ 2019-06-23 16:05  逆天而行。  阅读(516)  评论(0编辑  收藏