Reproduce “One Adapter, Many Tasks: Task-Conditioned Feature Transformations for Continual 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
Yunxiang Fu, Meng Lou, Yizhou Yu
FACET enables exemplar-free class-incremental learning with a frozen pretrained vision backbone as new classes arrive without replaying old examples. It shares one adapter whose feature transformation depends on predicted task context, and uses a consistency objective to limit drift. On OmniBenchmark-1K with 200 tasks, it reaches 65.95% final accuracy versus MOS at 63.51%; the result is bounded to this setting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.