Wasserstein Proximal of GANs

TL;DR

This paper presents a novel Wasserstein-2 metric proximal method for training GANs, leveraging Wasserstein information geometry to improve training speed and stability.

Contribution

It introduces a parametrization invariant natural gradient approach using Wasserstein geometry, providing easy-to-implement regularizers for implicit deep generative models.

Findings

Enhanced training speed and stability

Reduced Fréchet Inception Distance

Improved wall-clock time performance

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 Fr\'echet Inception Distance.