e-means: outperforming k-means (in one-tenth the epochs)
The authors report 10–45× GPU speedups while reaching lower clustering loss than Lloyd’s k-means.
E-means updates cluster centers using an exponential moving average with mini-batches. The paper derives rules for setting decay rate and batch size without hyperparameter sweeps. Across embedding datasets, it beats Lloyd’s 25-epoch loss in 1.4–2.6 epochs. In one example, it groups 10 billion DINO embeddings into 16 million clusters in 2.5 epochs, again finishing below Lloyd’s 25-epoch loss.
HF Daily Papers' note
E-means updates cluster centers using an exponential moving average with mini-batches. The paper derives rules for setting decay rate and batch size without hyperparameter sweeps. Across embedding datasets, it beats Lloyd’s 25-epoch loss in 1.4–2.6 epochs. In one example, it groups 10 billion DINO embeddings into 16 million clusters in 2.5 epochs, again finishing below Lloyd’s 25-epoch loss.
HF Daily Papers' note
score 5