GPU implementation of the FastICA algorithm
Gil Benkö, Zoltán Juhász · 2019
Independent Component Analysis is a widely used method in EEG data processing for removing unwanted artefacts from the measured data. The drawback of this method is its high computational cost, resulting in long execution times. A massively parallel GPU implementation of the popular FastICA algorithm is presented in this paper. The implementation uses standard CUDA library functions where possible and custom parallel kernels for the remaining steps. The results show that for typical EEG processing setups our version can be executed within real-time limits, allowing sophisticated automatic artefact removal algorithms to be executed during measurement.