Reproduce “Long-Tailed 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
Xialei Liu, Yu-Song Hu, Xu-Sheng Cao, Andrew D. Bagdanov, Ke Li, Ming-Ming Cheng
The authors adapt two-stage classifier balancing to class-incremental learning without task boundaries at inference. They define ordered and shuffled long-tailed settings, then train a learnable weight-scaling layer with balanced sampling after the main learning stage, freezing earlier task heads. On ImageNet-Subset, shuffled LT-CIL, five-task EEIL, average incremental accuracy rises from 42.87% to 49.31%; gains vary across tested settings.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.