On Catastrophic Forgetting and Mode Collapse in Generative Adversarial Networks

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.

arXiv commentThis is an extended version of our paper in ICML'18 Workshop on Theoretical Foundation and Applications of Deep Generative Models. Accepted to IJCNN 2020

References in corpus (12)

• Overcoming catastrophic forgetting in neural networks
• GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium
• Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)
• Continual Learning Through Synaptic Intelligence
• Which Training Methods for GANs do actually Converge?
• On Convergence and Stability of GANs
• Generalization and Equilibrium in Generative Adversarial Nets (GANs)
• The Numerics of GANs
• Continual Learning in Generative Adversarial Nets
• Improving Generalization and Stability of Generative Adversarial Networks
• On the Discrimination-Generalization Tradeoff in GANs
• Generative Adversarial Network Training is a Continual Learning Problem

Cited by in corpus (7)

• Task Agnostic Continual Learning Using Online Variational Bayes with Fixed-Point Updates
• Self-Supervised GANs via Auxiliary Rotation Loss
• Autoencoding Generative Adversarial Networks
• Learning C to x86 Translation: An Experiment in Neural Compilation
• Frosting Weights for Better Continual Training
• Application of Deep Learning-based Interpolation Methods to Nearshore Bathymetry
• Toward a Generalization Metric for Deep Generative Models