An Efficient Method for Incremental Learning of GMM Using CUDA

Abstract

Incremental learning algorithms of the Gaussian Mixture Model can find applications in various scenarios. This paper proposes a CUDA-based method to accelerate incremental learning of GMM. Different from existing methods towards GMM on GPU, our method aims to hide data transfer latency instead of accelerating the algorithm itself. Due to the inherent characteristic of memory-critical incremental learning applications, loading data from external memory and copying data from host to device will inevitably contributes to the overall time consumption. CUDA capabilities called "concurrent execution" and "overlap data transfer" are leveraged to implement incremental GMM learning in a pipelined pattern. The efficiency of our method is validated through preliminary experiments, which demonstrate improved performance over the non-pipelined method.