Stabilize continual learning with hyperspherical replay
Keywords: Continual learning
TL;DR: This paper proposed to reduce the stability gap in continual learning with adaptive angular replay.
Abstract: Neural networks face catastrophic forgetting of previously learned knowledge
when training on new task data. While the field of continual learning has made
promising progress in reducing this forgetting, recent work has uncovered an
interesting phenomenon: existing techniques often exhibit a sharp performance
drop on prior tasks during the initial stages of new task training, a phenomenon
known as the ”stability gap.” This phenomenon not only raises safety concerns
but also challenges the current understanding of neural network behavior in continual learning scenarios. Inspired by this discovery, we revisit two fundamental
questions in continual learning: 1) Is the past learned knowledge within deep
networks lost abruptly or gradually? and 2) Is past learned knowledge ever completely erased? Our analysis reveals that abrupt forgetting occurs not only in the
final fully connected layer but also permeates the feature space and most layers,
sparing only the earliest layers. Alarmingly, a single gradient update can severely
disrupt the learned class structure. We identify degenerate solutions in the softmax
cross-entropy loss as a major contributing factor, with memory samples exhibiting
higher feature norms compared to new samples. To address these issues, we pro-
pose Adaptive Angular Replay (AAR), a simple yet effective approach that learns
features in hyperspherical space using feature and weight normalization. Angular
ER demonstrates a strong ability to preserve class structure during task transitions. Additionally, we introduce an adaptive scaling strategy to further mitigate
the stability gap and improve overall accuracy.
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 12919
Loading