Speedup of Fuzzy Clustering Through Stream Processing on Graphics Processing Units

摘要

As the number of data points, feature dimensionality, and number of centers for clustering algorithms increase, computational tractability becomes a problem. The fuzzy c-means has a large degree of inherent algorithmic parallelism that modern CPU architectures do not exploit. Many pattern recognition algorithms can be sped up on a graphics processing unit (GPU) as long as the majority of computation at various stages and the components are not dependent on each other. We present a generalized method for offloading fuzzy clustering to a GPU, while maintaining control over the number of data points, feature dimensionality, and the number of cluster centers. GPU-based clustering is a high-performance low-cost solution that frees up the CPU. Our results show a speed increase of over two orders of magnitude for particular clustering configurations and platforms.