Implementing Independent Component Analysis in General-Purpose GPU Architectures

Abstract

New computational architectures, such as multi-core processors and graphics processing units (GPUs), pose challenges to application developers. Although in the case of general-purpose GPU programming, environments and toolkits such as CUDA and OpenCL have simplified application development, different ways of thinking about memory access, storage, and program execution are required. This paper presents a strategy for implementing a specific signal processing technique for blind-source separation: infomax independent component analysis (ICA). Common linear algebra operations are mapped to a low cost programmable graphics card using the OpenCL programming toolkit. Because many components of ICA are inherently parallel, ICA computations can be accelerated by low cost parallel hardware. Experimental results on simulated and speech signals indicate that efficiency gains and scalability are achievable through general-purpose GPU implementation, and suggest that important applications in telecommunications, speech processing, and biomedical signal analysis can benefit from these new architectures. The utilization of low cost GPUs for programming may potentially facilitate real-time applications of previously offline algorithms.