Plasticity by Precision: Exemplar-free Analytic Adaptation for Class-Incremental Learning

TMLR Paper8737 Authors

Abstract: Class-Incremental Learning (CIL) aims to enable models to acquire new knowledge sequentially while preserving previously learned information, emulating human-like learning capabilities. Current methods, including pre-trained foundation models and Experience Replay (ER) methods, serve as strong baselines for sequential task learning. However, these methods remain prone to catastrophic forgetting, especially in online settings with non-stationary data and blurry task boundaries. Additionally, the requirement to store historical samples in ER-based methods introduces significant memory overhead and privacy risks, limiting the practical adoption of CIL models in real-world applications. To address this, we propose an \textbf{E}xemplar-\textbf{f}ree \textbf{A}nalytic \textbf{A}daptation for \textbf{C}lass-\textbf{I}ncremental \textbf{L}earning (AACL) framework that updates the classifier in a principled probabilistic manner. Our key contribution is a closed-form Bayesian update that unifies three critical components: (1) the prior precision encapsulating knowledge from previous tasks, (2) a Fisher Information-inspired weight penalty to protect learned knowledge, and (3) the feature correlation matrix representing evidence from new data. Our framework balances plasticity and stability by integrating prior knowledge with streaming data, preserving learned representations while adapting to new tasks. We conduct comprehensive evaluations on benchmark datasets under the SI-Blurry setting, achieving $\mathcal A_{\text{AUC}}$ improvements of 8\%, 3\%, and 4\% on CIFAR-100, ImageNet-R, and Tiny-ImageNet, respectively.

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

Assigned Action Editor: ~Robert_M._Gower1

Submission Number: 8737