PSAG: Projection-based Stabilized Attribution Guidance for Online Continual Learning

TMLR Paper6802 Authors

Abstract: Online Continual Learning (OCL) aims to incrementally learn from non-stationary data streams in a one-pass setting, facing the dual challenges of catastrophic forgetting and insufficient training. These challenges intensify the stability-plasticity dilemma, where preserving old knowledge conflicts with acquiring new information. In this paper, we propose Projection-based Stabilized Attribution Guidance (PSAG), a modular framework that leverages gradient-based attributions as active guidance signals to selectively preserve task-relevant representations. Our framework consists of three complementary mechanisms: (1) Attribution-Guided Feature Modulation (AGFM) that anchors critical features in the representation space; (2) Importance-Aware Loss Reweighting (IALR) that prioritizes informative samples at the loss level; and (3) Manifold-Consistent Projection (MCP) that emphasizes critical feature dimensions within a Riemannian metric space. To address the issue of attribution instability in online continual learning, we introduce a {Reliable Reference Model (R-Model)} that maintains consistent knowledge through exponential moving average updates. This design prevents feedback loops during attribution computation and enables reliable feature importance estimation. Extensive experiments on Split CIFAR-10, Split CIFAR-100, and Split Mini-ImageNet demonstrate that PSAG achieves consistent improvements over strong baselines, confirming the efficacy of stabilized attribution guidance in resolving the stability-plasticity dilemma.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Xavier_Alameda-Pineda1

Submission Number: 6802