Application of Dictionary Learning in Alleviating Computational Burden of EEG Source Localization

Abstract

Two techniques are proposed to alleviate the computational burden of MUltiple SIgnal Classification (MUSIC) algorithm applied to Electroencephalogram (EEG) source localization. A significant reduction was achieved by parsing the cortex surface into smaller regions and nominating only a few regions for the exhaustive search inherent in the MUSIC algorithm. The nomination procedure involves a dictionary learning phase in which each region is assigned an atom matrix. Moreover, a dimensionality reduction step provided by excluding some of the electrodes is designed such that the Cramer-Rao bound of localization is maintained. It is shown by simulation that computational complexity of the MUSIC-based localization can be reduced by up to $80\%$.