A GPU Implementation of FastICA in Audio Applications for Small Number of Components

Extracting independent components from audio data has plenty of uses in biology, music, communication and media and in many other fields. The FastICA algorithm is a relatively fast and simple algorithm, that assumes the original sources to be nongaussian and works by lowering the gaussianity of the mixed sources. Yet as the number of samples increase so does the time required for its execution. While one solution would be to simply use just a subset of the samples, in this paper we look at the possibility of extending the FastICA algorithm to the GPU. While similar efforts have been pursued in the past, we deal with data of relatively few components, as would be more common when dealing with audio data. We implement a fully GPU FastICA as well as a CPU-GPU hybrid algorithm, both based on the CUDA platform and compare them with the CPU version. Our results indicate that for large samples the CPU-GPU hybrid and the GPU algorithms perform better than their CPU counterpart.