Classification of EEG Signals Using Blind Source Separation Techniques and Deep Networks
The objective of this work is to study blind source separation techniques based on second-order statistics (SOS) applied to electroencephalography (EEG) signals, in order to analyze their effect on the performance of deep classifiers. The motor imagery paradigm is considered, and the preprocessing step is done using blind source separation techniques assuming the Post-Nonlinear mixtures context. As the objective is to focus on methods based on SOS for extracting independent components, the classic AMUSE and SOBI algorithms will also be used for a comparative analysis. The independent components recovered are then classified using EEGNet, a deep neural network. The performance of blind source separation preprocessing techniques will be evaluated through the use of deep networks that currently are the state-of-the-art methods in this scope.