On catastrophic forgetting and mode collapse in Generative Adversarial Networks

By Hoang Thanh-Tung, Truyen Tran, Svetha Venkatesh. 2018

64 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Statistical and Nonlinear Physics, Generative Models

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Chunyuan Li
2. Lawrence Carin
3. Neil Houlsby
4. Xiaohua Zhai
5. Jianfeng Gao
6. Ting-Li Chen
7. Ngai-Man Cheung
8. Ngoc-Trung Tran

In Computer Vision and Pattern Recognition

1. Luc Van Gool
2. Dacheng Tao
3. Chunyuan Li
4. Trevor Darrell
5. Jitendra Malik
6. Xiaogang Wang
7. Ming-Hsuan Yang
8. Cordelia Schmid

In Statistical and Nonlinear Physics

1. Ting-Li Chen
2. Daniel Soudry
3. Elad Hoffer
4. Ngai-Man Cheung
5. Kevin J Liang
6. Lawrence Carin
7. Viet-Hung Tran
8. Ngoc-Bao Nguyen

In Generative Models

1. Ngai-Man Cheung
2. Ngoc-Trung Tran
3. Xiaohua Zhai
4. Lawrence Carin
5. Chunyuan Li
6. Ting-Li Chen
7. Kevin J Liang
8. Mario Lucic

Citations in the corpus

• Self-Supervised GANs via Auxiliary Rotation Loss (2018, background): “[14] study catastrophic forgetting in the discriminator and mode collapse, relating these to training instability.”
• Task-Agnostic Continual Learning Using Online Variational Bayes With Fixed-Point Updates (2020, methodology): “Another direction is to use FOO-VB to improve GAN stability (Thanh-Tung & Tran, 2018).”
• Improving GAN with neighbors embedding and gradient matching (2018, unknown): “In addition, we follow (Thanh-Tung, Tran, and Venkatesh 2018) to explore the gradient map of the discriminator scores of compared methods: standard GAN, WGAN-GP, Dist-GAN and ours as shown in Fig.”
• Generative Adversarial Network Training is a Continual Learning Problem (2018, background): “Thanh-Tung et al. (2018) independently reached a similar conclusion on catastrophic forgetting in GANs, but focused on gradient penalties and momentum on toy problems.”
• On Data Augmentation for GAN Training (2020, background): “This constraint mitigates gradient vanishing [20] and catastrophic forgetting [27].”