Reproduce “e-means: outperforming k-means (in one-tenth the epochs)” 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
Swann Bessa, Pierre-Emmanuel Mazaré, Gergely Szilvasy, Hugo Touvron, David Picard, Matthijs Douze, Hervé Jégou
e-means clusters fixed vector collections by updating centroids from batches using exponential moving averages, with batch size and update rate set by scaling rules. On a 1-billion-vector DINO10B subset with 1,048,576 centroids, it reaches Flash-KMeans’ held-out test MSE in 1.6 epochs, with a reported 45.5× GPU speedup. These scaling rules are not optimal for very low-dimensional, tiny problems.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.