Reproduce “SCAN: Learning to Classify Images without Labels” 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

Wouter Van Gansbeke, Simon Vandenhende, Stamatios Georgoulis, Marc Proesmans, Luc Van Gool

SCAN clusters images by learning features through instance discrimination, then using nearest neighbors to train a clustering model and refine predictions through self-labeling. On CIFAR-10, it reaches validation clustering accuracy of 87.6 ± 0.4, versus 65.9 ± 5.7 for K-means on pretext features. Training uses no labels, but sets cluster count to the known class count; accuracy is scored against ground-truth labels.

Summary by alphaXiv.

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