Skip to content
NewA calmer, faster ZeroHour
StoryAI researchinfo

Authors say e-means beats Lloyd k-means in far fewer epochs

2 sources · 2 articlesFirst reported Updated ()
What's new

November 9: Hacker News adds Flash-KMeans and Faiss comparisons, a cosine schedule, dead-cluster revival, finalization, and a wider 1.2–3 epoch range.

Merged summary · x-ai/grok-4.7 · rewritten as coverage arrivesSummaries by Vani AI

Authors say e-means, an EMA mini-batch k-means method, beats Lloyd loss in a few epochs with large GPU speedups.

Authors present e-means, mini-batch k-means updated with an exponential moving average and deep-learning tooling. They derive scaling laws and a formula for decay rate and batch size so those settings need no hyperparameter sweep. Hugging Face daily papers reports a lower loss than Lloyd k-means at 25 epochs in 1.4 to 2.6 epochs; Hacker News reports 1.2 to 3 epochs. Both report 10–45× GPU speedups; Hacker News attributes the GPU comparison to Flash-KMeans and also cites 8–15× CPU speedups against Faiss. On DINO embeddings, the authors say 10 billion points were clustered into 16 million clusters in 2.5 epochs, often beating k-means++.

Key facts
  • e-means uses exponential moving average mini-batch centroid updates.
  • Decay rate and batch size come from a no-sweep scaling-law formula.
  • Hugging Face: 1.4–2.6 epochs; Hacker News: 1.2–3 epochs to beat Lloyd at 25.
  • Authors report 10–45× GPU speedups; Hacker News also cites 8–15× versus Faiss on CPU.
  • Authors say DINO10B formed 16 million clusters in 2.5 epochs.
  • Hacker News describes a cosine anneal, dead-cluster revival, and exact-mean finalization.
Explain this AI news with Vani AIWhat was announced, what is new and why it matters, in a new chat

What each outlet adds

beyond the shared facts
  • Hugging Face daily papersDINO10B result, k-means++ comparison, and 1.4–2.6 epoch range.
  • Hacker NewsFlash-KMeans and Faiss baselines, schedule, revival, and finalization.

Coverage

oldest first · one row per article · UTC
  1. · 5d ago Hugging Face daily papersFirst report

    Paper introduces e-means, a faster mini-batch k-means variant

    Authors report 10-45x GPU speedups and a no-sweep rule for decay rate and batch size.

  2. · 1d ago Hacker News · AI

    E-means claims much faster k-means at equal or better loss

    Authors report beating Lloyd loss in 1.2 to 3 epochs and 10–45× GPU speedups versus Flash-KMeans.