A Method to Accelerate K-Means and GMM Computation with GPU and Multi-Core CPU

Abstract

K-means and Gaussian mixture model (GMM) clustering, as dictionary learning procedures, lie at the heart of BoW (BoW) framework. With the data scale increasing, it urgently requires efficient ways to realize these processes. In this paper, we present some new approaches to calculate k-means, GMM and MAP algorithms, which can be effectively accelerated by GPU, multicore CPU. The speed-up is empowered by matrix-based operations, and we show that these three procedures can be concisely reformulated into matrix multiplications, which can be efficiently accelerated by parallel computation facilities on single machine. In the experiments on music genre and mood classification, we show that compared to single threaded CPU execution, our approaches can achieve the acceleration for dictionary learning by38.0 and 209.5 for k-means and GMM clustering respectively, but with just less 1% performance decline.(Abstract)