2017.4-Jeff Donahue, Trevor Darrell-Adversarial feature learning-UCB-ICLR2017 阅读笔记

2017.4-Jeff Donahue, Trevor Darrell-Adversarial feature learning-UCB-ICLR2017

• 本文创新点：提出 BiGANs，能够进行 inverse mapping (data => latent space)
  • 在GAN 中引入 encoder，命名为 Bi-GAN, 将 discriminator 对 X 和 G(z) 的判别转化为对 (x, E(x)) 和 (G(z), z) 的判别。
  • 推导证明了 BiGAN 最优的 E 和 G 是互逆的： $E=G^{-1}$
  • 推导证明了 BiGAN 与 ${\ell }_0$ loss 下的 autoencoder 是 closely related.
• 本文的 traning, hyper param. setting 与 evaluation 等，均沿用前人文献。

Abstract

• Target of the research
  • Learn feature repre. for auxiliary problems where semantics are relevant.
• GAN
  • cannot project: data => latent space
• This paper
  • propose BiGANs
  • ability of inverse mapping: data => latent space
• Result
  • the resulting learned feature representation is useful for auxiliary supervised discrimination tasks

adversarial (minimax) objective

• GAN: ${\min }_G{\max }_{D}V(D,G)$

• BiGAN: ${\min }_{G,E}{\max }_{D}V(D,E,G)$
  • optimize the Jensen-Shannon divergence between the joint dist. $P_{EX}$ and $P_{GZ}$