On Catastrophic Forgetting and Mode Collapse in Generative Adversarial Networks

TL;DR

This paper reveals that GANs experience catastrophic forgetting during training, which relates to mode collapse and non-convergence, and explores methods to mitigate this issue.

Contribution

It conceptualizes GAN training as a continual learning problem and analyzes the discriminator's landscape to understand and address catastrophic forgetting.

Findings

GANs suffer from catastrophic forgetting even for a single target distribution.

The level of task mismatch influences the extent of forgetting.

Preventing forgetting can improve GAN convergence and reduce mode collapse.

Abstract

In this paper, we show that Generative Adversarial Networks (GANs) suffer from catastrophic forgetting even when they are trained to approximate a single target distribution. We show that GAN training is a continual learning problem in which the sequence of changing model distributions is the sequence of tasks to the discriminator. The level of mismatch between tasks in the sequence determines the level of forgetting. Catastrophic forgetting is interrelated to mode collapse and can make the training of GANs non-convergent. We investigate the landscape of the discriminator's output in different variants of GANs and find that when a GAN converges to a good equilibrium, real training datapoints are wide local maxima of the discriminator. We empirically show the relationship between the sharpness of local maxima and mode collapse and generalization in GANs. We show how catastrophic forgetting prevents the discriminator from making real datapoints local maxima, and thus causes non-convergence. Finally, we study methods for preventing catastrophic forgetting in GANs.