Reproduce “Adaptive Confidence Smoothing for Generalized Zero-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
Yuval Atzmon, Gal Chechik
COSMO addresses generalized zero-shot learning, where classifiers recognize seen classes and classes known through descriptions. It keeps a seen-class expert and a zero-shot expert separate, using a gate trained on held-out seen classes as proxies to estimate familiarity, soften unfamiliar predictions, and combine outputs. On AWA, COSMO+LAGO reaches 63.6% harmonic-mean accuracy versus 61.5% for SE-GZSL, but underperforms cycle-(U)WGAN on CUB.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.