Revisiting Prototype Rehearsal for Exemplar-Free Continual Learning

Manifold-aware boundary sampling with adaptive class-balanced loss

Prototype rehearsal overview

This project revisits prototype rehearsal for exemplar-free class-incremental learning. We argue that prototype rehearsal is not fundamentally limited; rather, common implementations suffer from two issues: they treat prototypes as isolated class summaries and they ignore the hidden imbalance between a few synthetic old-class samples and many real current-task samples.

To address these issues, we introduce Constrained Expansive Over-Sampling (CEOS) and Adaptive Class-Balanced (ACB) Loss. CEOS generates boundary-aware synthetic features by interpolating old-class prototypes toward nearest enemy features while preserving class separation. ACB Loss dynamically reweights classes over time to reduce the temporal imbalance between prototype-based old-class supervision and current-task data.

Highlights

  • Accepted to CVPR 2026 Findings.
  • Restores the competitiveness of prototype rehearsal in exemplar-free CIL.
  • Combines manifold-aware boundary sampling with time-aware class balancing.
  • Code: HXuSz11/ACB_CEOS_CVPR2026_Findings