E-means: 10x faster k-means at equal loss (e-means.github.io)

A groundbreaking algorithm called e-means has been introduced, achieving k-means clustering 10–45 times faster than the existing state-of-the-art methods like Flash-KMeans on GPU, and 8–15 times faster on CPU while showing improved accuracy. This advancement is notable as traditional k-means implementations often face a trade-off between speed and clustering quality; e-means overcomes this barrier by utilizing mini-batch updates with exponential moving averages that dynamically set batch and step sizes, eliminating the need for hyperparameter tuning. The significance of e-means lies in its ability to handle dead clusters effectively by adopting a proportional splitting strategy, thus enhancing codebook capacity, which is increasingly vital for large datasets. By including a cosine-scheduled decay rate and a finalization stage to ensure optimal loss reduction, e-means not only accelerates the clustering process but also improves convergence rates and reliability in escaping local minima compared to Lloyd’s algorithm. This novel approach has been validated across diverse datasets and demonstrates robust generalization, making it a promising tool for the AI/ML community engaged in data clustering tasks.