Reproduce “ScaIL: Classifier Weights Scaling for 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

Eden Belouadah, Adrian Popescu

ScaIL addresses class-incremental visual classification by retaining each class’s initial classifier weights and rescaling them for later updates, while using bounded image exemplars and the current feature extractor. On four datasets with ResNet-18, it scores −3.71 with herding on the aggregate normalized top-5 accuracy-gap measure, versus −6.40 for vanilla fine-tuning. Results cover tested memory configurations; BiC does better at 10 incremental states.

Summary by alphaXiv.

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