Degradation of Feature Space in Continual Learning

TL;DR

This paper investigates whether promoting isotropic feature spaces improves continual learning, finding that such regularization often degrades performance due to fundamental differences in feature geometry compared to centralized training.

Contribution

The study reveals that isotropic regularization does not enhance and can harm continual learning performance, highlighting the need to reconsider feature geometry assumptions in non-stationary environments.

Findings

Isotropic regularization fails to improve continual learning accuracy.

Feature space in continual learning tends to become anisotropic.

Isotropy beneficial in centralized training may not apply to continual learning.

Abstract

Centralized training is the standard paradigm in deep learning, enabling models to learn from a unified dataset in a single location. In such setup, isotropic feature distributions naturally arise as a mean to support well-structured and generalizable representations. In contrast, continual learning operates on streaming and non-stationary data, and trains models incrementally, inherently facing the well-known plasticity-stability dilemma. In such settings, learning dynamics tends to yield increasingly anisotropic feature space. This arises a fundamental question: should isotropy be enforced to achieve a better balance between stability and plasticity, and thereby mitigate catastrophic forgetting? In this paper, we investigate whether promoting feature-space isotropy can enhance representation quality in continual learning. Through experiments using contrastive continual learning techniques on CIFAR-10 and CIFAR-100 data, we find that isotropic regularization fails to improve, and can in fact degrade, model accuracy in continual settings. Our results highlight essential differences in feature geometry between centralized and continual learning, suggesting that isotropy, while beneficial in centralized setups, may not constitute an appropriate inductive bias for non-stationary learning scenarios.