Reproduce “Subclass Distillation” 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
Rafael Müller, Simon Kornblith, Geoffrey Hinton
Subclass distillation trains a student to match a teacher’s probabilities over invented subdivisions within each class, rather than only probabilities over the classes, without new labels. On CIFAR-2x5, an AlexNet student reached 91.4% binary accuracy versus 90.1% with penultimate-layer distillation. On Criteo, it learned faster in steps and improved validation accuracy with 10% transfer data; full-data final accuracies were comparable.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.