E-means: Outperforming k-means in one-tenth the epochs
The gist
Meta FAIR open-sourced e-means, which reaches Lloyd's 25-epoch loss in 1.2 to 3 epochs.
Faster clustering at equal loss could cut training costs across ML pipelines.
How it unfolded
- Nov 9, 2026e-means is open-sourced and posted as arXiv:2611.09501, titled "e-means: outperforming k-means (in one-tenth the epochs)", by Bessa, Mazaré, Szilvasy, Touvron, Picard, Douze and Jégou.
- Nov 9, 2026The authors report e-means reaches Lloyd's 25-epoch test loss in 1.2 to 3 epochs, a 10–45× wall-clock reduction on GPU against Flash-KMeans and 8–15× on CPU against Faiss.
- Nov 9, 2026On 11 VIBE embedding datasets it needs 1.6 to 2.4 epochs to match Lloyd's 25-epoch test error; at billion scale on DINO10B, 10 billion image-patch vectors are clustered into about 16 million clusters in 2.5 epochs.
Key terms
- k-means
- A clustering method that partitions observations into k groups, each assigned to the nearest cluster mean.
- Lloyd's algorithm
- An iterative algorithm that repeatedly finds each partition's centroid and reassigns points to the nearest centroid.
- exponential moving averages
- A moving average is a statistical calculation that averages data points over successive overlapping periods to smooth out short-term fluctuations.
Source: Wikipedia