GPU optimization of electroencephalogram analysis

Abstract

Nowadays, with the advent of new non-invasive techniques of brain imaging, researchers have access to neural processes underlying the cognition in humans. One of the main challenges in this techniques is the detection of patterns in brain signals, generally very noisy and with artifacts inserted by vital signs. One of the most successful techniques for this is Independent Component Analysis which detects statistically independent components that are produced from different sources. These methods are very expensive in computational time, with many hours of processing for a single experiment. We analyzed this algorithm and detect two main types of operations: vector-matrix and matrix-matrix. We implemented an ad-hoc solution that executes on GPU and compared this with the original and CUBLAS versions. We obtained a 4x and 40x of performance increase of vector-matrix and matrix-matrix operations, respectively. These results are the first step towards real-time EEG processing which may produce a significant advance into BCI applications.