Prescribed Generative Adversarial Networks

  • 2019-10-09 23:40:05
  • Adji B. Dieng, Francisco J. R. Ruiz, David M. Blei, Michalis K. Titsias
  • 96

Abstract

Generative adversarial networks (GANs) are a powerful approach tounsupervised learning. They have achieved state-of-the-art performance in theimage domain. However, GANs are limited in two ways. They often learndistributions with low support---a phenomenon known as mode collapse---and theydo not guarantee the existence of a probability density, which makes evaluatinggeneralization using predictive log-likelihood impossible. In this paper, wedevelop the prescribed GAN (PresGAN) to address these shortcomings. PresGANsadd noise to the output of a density network and optimize anentropy-regularized adversarial loss. The added noise renders tractableapproximations of the predictive log-likelihood and stabilizes the trainingprocedure. The entropy regularizer encourages PresGANs to capture all the modesof the data distribution. Fitting PresGANs involves computing the intractablegradients of the entropy regularization term; PresGANs sidestep thisintractability using unbiased stochastic estimates. We evaluate PresGANs onseveral datasets and found they mitigate mode collapse and generate sampleswith high perceptual quality. We further found that PresGANs reduce the gap inperformance in terms of predictive log-likelihood between traditional GANs andvariational autoencoders (VAEs).

 

Quick Read (beta)

loading the full paper ...