E-means claims much faster k-means at equal or better loss
Original title: E-means: 10x faster k-means at equal loss
Authors report beating Lloyd loss in 1.2 to 3 epochs and 10–45× GPU speedups versus Flash-KMeans.
The e-means project proposes mini-batch k-means with exponential moving averages, scaling-law hyperparameters, and fused GPU kernels. Across reported scales, it reaches a test loss below Lloyd's at 25 epochs in 1.2 to 3 epochs. Measured end to end, the authors claim 10–45× GPU speedups against Flash-KMeans and 8–15× on CPU against Faiss. They say EMA stays inside the convex hull of assigned points, unlike large-step gradient descent.
Unconfirmed: Speed and loss figures are the project's own measurements, not an independent evaluation.
- Batch size and step size are derived from scaling laws, with no manual tuning.
- A cosine schedule anneals the step from alpha0 to 0.1 alpha0.
- Dead clusters below a size threshold are revived by splitting a large cluster.
- A finalization stage sets each centroid to its exact last-epoch mean.
From the original article
11 min read at e-means.github.io
Across a large range of scales and datasets , e-means reaches a test loss below Lloyd's at 25 epochs, in 1. 2 to 3 epochs. Measured end to end, that is a 10–45× wall-clock reduction on GPU against the state-of-the-art Lloyd implementation (Flash-KMeans [2] ), and 8–15× on CPU against Faiss [1] .
By porridgeraisin · e-means.github.io
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.