Reducing Noise in GAN Training with Variance Reduced Extragradient

TL;DR

This paper introduces SVRE, a variance-reduced extragradient method that mitigates stochastic gradient noise in GAN training, leading to more stable convergence and improved efficiency over traditional methods.

Contribution

The paper proposes a novel stochastic variance-reduced extragradient algorithm that enhances convergence and stability in GAN training compared to existing approaches.

Findings

SVRE performs comparably to batch methods on MNIST with lower computational cost.

SVRE yields more stable GAN training on standard datasets.

The method improves convergence rates for a broad class of game optimization problems.

Abstract

We study the effect of the stochastic gradient noise on the training of generative adversarial networks (GANs) and show that it can prevent the convergence of standard game optimization methods, while the batch version converges. We address this issue with a novel stochastic variance-reduced extragradient (SVRE) optimization algorithm, which for a large class of games improves upon the previous convergence rates proposed in the literature. We observe empirically that SVRE performs similarly to a batch method on MNIST while being computationally cheaper, and that SVRE yields more stable GAN training on standard datasets.