e-means: 10x faster k-means at equal loss
Web page · https://e-means.github.io/
First seen by Oriel on .
AI summary
e-means is a mini-batch k-means method using exponential moving averages that beats 25-epoch Lloyd test loss in 1.2–3 epochs, delivering 10–45× GPU and 8–15× CPU speedups.
- It automatically derives batch size and initial step size from scaling laws over n, k, and T, then applies cosine scheduling and exact last-epoch finalization.
- Its exponential moving averages forget stale assignments, remain stable at large steps, and combine with fused GPU kernels to convert epoch reductions into wall-clock gains.
- Dead clusters are revived by splitting large active clusters sampled proportional to size, preserving effective codebook capacity better than replacement with random batch samples.
- Across four toy problems and 200 starts, e-means escaped pathological local minima substantially more often than Lloyd, with two cases reducing failure rates to 0%.
- Emeans and NumpyEmeans are available in faiss/contrib/e_means, with five training epochs by default and support for PyTorch tensors or NumPy.
#k-means #clustering #faiss #gpu computing
Generated by gpt-5.6-sol. Check the original for details.
Metric trends
Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained
Points
Score reported by the source.
- Hacker News: 501, latest observation ; +487 over observed period since
- Lobsters: 3, latest observation ; +2 over observed period since
Comments
Comment count reported by the source.
- Hacker News: 78, latest observation ; +78 over observed period since
- Lobsters: 0, latest observation ; 0 over observed period since
Aliases
| Source | External ID | URL |
|---|---|---|
| Hacker News | 50429734 | https://e-means.github.io/ Read discussion |
| Lobsters | rnmylu | https://e-means.github.io/ Read discussion |