Reproduce “Concept Learners for 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
Kaidi Cao, Maria Brbic, Jure Leskovec
COMET makes few-shot classification compare examples through named, reusable concepts rather than one shared similarity space. It embeds masked input dimensions, forms class prototypes for each concept, and combines distances to predict a label. On Tabula Muris cross-organ cell-type classification, its 1-shot 5-way accuracy is 79.4 ± 0.9%, versus 67.2 ± 0.8% for ProtoNetEns, a result limited to the reported task and conditions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.