Wasserstein Proximal of GANs

2021/02/13 by Lin, Alex Tong, Li, Wuchen, Osher, Stanley +1 · 3 citations
#Artificial Intelligence (cs.AI) #FOS: Computer and information sciences #FOS: Mathematics #Machine Learning (cs.LG) #Numerical Analysis (math.NA)

paper · doi:10.48550/arxiv.2102.06862

Abstract

We introduce a new method for training generative adversarial networks by applying the Wasserstein-2 metric proximal on the generators. The approach is based on Wasserstein information geometry. It defines a parametrization invariant natural gradient by pulling back optimal transport structures from probability space to parameter space. We obtain easy-to-implement iterative regularizers for the parameter updates of implicit deep generative models. Our experiments demonstrate that this method improves the speed and stability of training in terms of wall-clock time and Fréchet Inception Distance.

Cited by

• An Optimal Transport-Based Method for Computing LM Rate and Its Convergence Analysis
• Expansive Natural Neural Gradient Flows for Energy Minimization
• Inexact JKO and proximal-gradient algorithms in the Wasserstein space

Related

• Training Wasserstein GANs without gradient penalties
• On the estimation of the Wasserstein distance in generative models
• A Differential Equation Approach for Wasserstein GANs and Beyond
• (q,p)-Wasserstein GANs: Comparing Ground Metrics for Wasserstein GANs
• Improved Training of Wasserstein GANs