BriefGPT.xyz
Nov, 2016
展开型生成对抗网络
Unrolled Generative Adversarial Networks
HTML
PDF
Luke Metz, Ben Poole, David Pfau, Jascha Sohl-Dickstein
TL;DR
我们引入了一种方法来稳定生成对抗网络（GAN），即通过将生成器目标定义为对鉴别器的展开优化来训练，从而解决了GAN训练中的常见问题和提高了生成器的多样性和数据分布覆盖率。
Abstract
We introduce a method to stabilize
generative adversarial networks
(GANs) by defining the generator objective with respect to an
unrolled optimization
of the
→