Meta researchers propose e-means, a faster upgrade to k-means
The method needs about a tenth of the training passes with lower error.

alphaXiv (@askalphaxiv) highlighted a paper from Meta researchers proposing e-means, which outperforms k-means with about 1/10 the training passes and lower clustering error.
It updates cluster centers with smaller batches and exponential moving averages, and uses scaling laws to pick batch size and learning rate. The post says it reaches lower error in 1.4-2.6 epochs versus 25 for k-means, with 10-45x GPU speedups, and can cluster 10 billion embeddings into 16 million groups in 2.5 epochs.
Why it mattersK-means is one of the most used algorithms in ML.
The post drew about 515 likes by Wednesday morning.




