Generalized and hybrid fast-ICA implementation using GPU
Titus Nanda Kumara, Hasindu Gamaarachchi, Geesara Kulathunga, Roshan Ragel · 2016
Independent Component Analysis is proposed as a solution to the Blind Source Separation problem. Among many of its realizations such as Infomax-ICA, Fast-ICA, and EASI- ICA, the Fast-ICA algorithm is the most famous and considered to be computationally the most efficient. Although the most capable, Fast-ICA still consumes a considerable amount of time on CPUs in real world implementations. Therefore, researchers in the past have considered accelerating Fast-ICA with the assistance of emerging GPGPU technologies. Such accelerations are effective because Fast-ICA makes use of a vast number of matrix operations which GPGPU are good at. However, all such previous works have focused on a particular application and therefore cannot be considered as generalized approaches for different data sizes. Therefore, in this paper, we concentrate on proposing a generalized and hybrid Fast-ICA implementation that makes use of the CPU and the GPGPU at their best possible capacities. We recommend an approach, where, depends on the data size, the different components of the Fast-ICA algorithm is suggested to be run on either CPU or GPGPU for the best performance. We achieve this via performing a deep analysis and profiling of the Fast-ICA algorithm on both the CPU and GPGPU and identifying the limitations and boundaries of the algorithm.