Reproduce “Adapter Merging with Centroid Prototype Mapping for Scalable Class-Incremental Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Takuma Fukuda, Hiroshi Kera, Kazuhiko Kawamoto

ACMap introduces an exemplar-free Class-Incremental Learning framework that integrates adapter merging with centroid prototype mapping. This approach achieves high accuracy while maintaining constant inference time, addressing a critical scalability limitation in current CIL methods by consolidating all task knowledge into a single, evolving adapter.

Key insights

  • Adapter merging with an initial weight replacement strategy allows for consolidating knowledge into a single adapter, enabling constant (O(1)) inference time while effectively mitigating catastrophic forgetting.
  • Centroid Prototype Mapping successfully realigns previously learned class representations in the evolving feature space, preventing semantic drift without requiring access to past data.
  • Leveraging large pre-trained Vision Transformers with parameter-efficient adapters provides a strong foundation for generalization and efficient adaptation in exemplar-free Class-Incremental Learning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.