COMPUTATIONAL INTELLIGENCE FOR BRAIN COMPUTER INTERFACE

Abstract

Substantial electroencephalography (EEG) research has been devoted to developing brain-computer interface. Nonetheless, several open issues remain. This thesis develops computational methodologies and contributes to EEG research from two aspects: the application of EEG in exoskeleton-assisted walking and the fundamental issues of EEG artifact processing. Firstly, we investigate cortical processes associated with locomotion assisted by a lower-limb exoskeleton. Spatio-spectral representation learning is proposed to decode EEG signals related to different amounts of assistive forces. Secondly, we investigate EEG artifacts that create formidable hurdles to all BCI applications. We design an experiment to characterize EEG artifacts and offer two heuristics for automated artifact removal. Lastly, we reformulate the artifact processing problem as an adversarial learning problem, and propose mitigative adversarial networks (MAN). MAN learns EEG artifact processing tasks end-to-end through back-propagation without artifact's type annotation, hand-crafted features, and loss functions. These works pave the way for practical and resilient EEG-based BCI.