e-means: 10x faster k-means at equal loss - Blog

e-means is a k-means clustering optimization method that utilizes exponential moving average updates, scaling laws for automatic hyperparameter tuning, and proportional split dead-cluster revival. Measured end-to-end, it achieves a 10–45x wall-clock reduction on GPU against Lloyd implementations and 8–15x on CPU against Faiss while reaching lower test loss. Open article →