BriefGPT.xyz
Jan, 2018
使用代表性特征改进生成式对抗网络的训练
Improved Training of Generative Adversarial Networks Using Representative Features
HTML
PDF
Duhyeon Bang, Hyunjung Shim
TL;DR
本论文使用预训练自编码器的代表特征隐式规范鉴别器，通过影响反向和正向KL距离使得GAN的训练更稳定，在图像质量和多样性上都有提高。
Abstract
Despite of the success of
generative adversarial networks
(
gans
) for image generation tasks, the trade-off between image diversity and visual quality are an well-known issue. Conventional techniques achieve eithe
→