NON-ODDBALL ERP PARADIGMS WITH JOINT TEMPORAL-FREQUENCY LEARNING IN CONVOLUTIONAL NEURAL NETWORK

Abstract

Brain-Computer Interface (BCI) helps people who have a severe disease to interact with external devices. Event-Related Potential (ERP) based BCI systems send stimuli, then detect brain signals responding to stimuli. Stimuli play loud or flashes with high intensity to distinguish the target brain signals in most researches. Thus, typical oddball-paradigm causes psychological and psychical discomforts. This thesis proposed non-oddball BCI paradigms which send stimuli which has almost zero volume or intensity. Users tend to produce voluntary mental task during experiment to cover typical stimuli’s brain signal, thus it compensates for the loss of accuracy of a result of the reduced stimulus. As an outcome, a specific mental task was used to investigate task-relevant endogenous components, and system performance was significantly enhanced. The proposed Convolutional Neural Network (CNN) approach has a decoding accuracy more than 90% for both the non-oddball visual and auditory paradigms, respectively, outperforming the linear classifier model noticeably. These discoveries offer new opportunities for pragmatic ERP systems, potentially improving the usability of current brain-computer interfaces remarkably.