Reproduce “Beyond Simple Meta-Learning: Multi-Purpose Models for Multi-Domain, Active and Continual Few-Shot 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

Peyman Bateni, Jarred Barber, Raghav Goyal, Vaden Masrani, Jan-Willem van de Meent, Leonid Sigal, Frank Wood

Simple CNAPS classifies few-shot image tasks with task-adapted features and support-derived class means and covariances, using Mahalanobis distance. Transductive CNAPS also uses unlabeled queries to adapt features and iteratively refine class parameters. Across 13 datasets, its average accuracy is 73.9%, versus 72.2% for Simple CNAPS. The authors report significant gains, but refinement can trail above 10 support examples per class.

Summary by alphaXiv.

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