Timezone: »

 
Poster
Retrospective Adversarial Replay for Continual Learning
Lilly Kumari · Shengjie Wang · Tianyi Zhou · Jeff A Bilmes

Tue Nov 29 09:00 AM -- 11:00 AM (PST) @ Hall J #121

Continual learning is an emerging research challenge in machine learning that addresses the problem where models quickly fit the most recently trained-on data but suffer from catastrophic forgetting of previous data due to distribution shifts --- it does this by maintaining a small historical replay buffer in replay-based methods. To avoid these problems, this paper proposes a method, ``Retrospective Adversarial Replay (RAR)'', that synthesizes adversarial samples near the forgetting boundary. RAR perturbs a buffered sample towards its nearest neighbor drawn from the current task in a latent representation space. By replaying such samples, we are able to refine the boundary between previous and current tasks, hence combating forgetting and reducing bias towards the current task. To mitigate the severity of a small replay buffer, we develop a novel MixUp-based strategy to increase replay variation by replaying mixed augmentations. Combined with RAR, this achieves a holistic framework that helps to alleviate catastrophic forgetting. We show that this excels on broadly-used benchmarks and outperforms other continual learning baselines especially when only a small buffer is available. We conduct a thorough ablation study over each key component as well as a hyperparameter sensitivity analysis to demonstrate the effectiveness and robustness of RAR.

Author Information

Lilly Kumari (University of Washington, Seattle)
Shengjie Wang (University of Washington)
Tianyi Zhou (University of Maryland, College Park)
Tianyi Zhou

Tianyi Zhou (https://tianyizhou.github.io) is a tenure-track assistant professor of computer science at the University of Maryland, College Park. He received his Ph.D. from the school of computer science & engineering at the University of Washington, Seattle. His research interests are in machine learning, optimization, and natural language processing (NLP). His recent works study curriculum learning that can combine high-level human learning strategies with model training dynamics to create a hybrid intelligence. The applications include semi/self-supervised learning, robust learning, reinforcement learning, meta-learning, ensemble learning, etc. He published >80 papers and is a recipient of the Best Student Paper Award at ICDM 2013 and the 2020 IEEE Computer Society TCSC Most Influential Paper Award.

Jeff A Bilmes (University of Washington, Seattle)

More from the Same Authors