Incorporation of non-euclidean distance metrics into fuzzy clustering on graphics processing units

Computational tractability of clustering algorithms becomes a problem as the number of data points, feature dimensionality, and number of clusters increase. Graphics Processing Units (GPUs) are low cost, high performance stream processing architectures used currently by the gaming, movie, and computer aided design industries. Fuzzy clustering is a pattern recognition algorithm that has a great amount of inherent parallelism that allows it to be sped up through stream processing on a GPU. We previously presented a method for offloading fuzzy clustering to a GPU, while maintaining full control over the various clustering parameters. In this work we extend that research and show how to incorporate non-Euclidean distance metrics. Our results show a speed increase of one to almost two orders of magnitude for particular cluster configurations. This methodology is particularly important for real time applications such as segmentation of video streams and high throughput problems.