Domain Adaptation with Source Selection for Motor-Imagery based BCI

Recent successes of deep learning methods in various applications have inspired BCI researchers for their use in EEG classification. However, data insufficiency and high intra- and inter-subject variabilities hinder from taking their advantage of discovering complex patterns inherent in data, which can be potentially useful to enhance EEG classification accuracy. In this paper, we devise a novel framework of training a deep network by adapting samples of other subjects as a means of domain adaptation. Assuming that there are EEG trials of motor-imagery tasks from multiple subjects available, we first select a subject whose EEG signal characteristics are similar to the target subject based on their power spectral density in resting-state EEG signals. We then use EEG signals of both the selected subject (called a source subject) and the target subject jointly in training a deep network. Rather than training a single path network, we adopt a multi-path network architecture, where the shared bottom layers are used to discover common features for both source and target subjects, while the upper layers branch out into (1) source-target subject identification, (2) label prediction optimized for a source subject, and (3) label prediction optimized for a target subject. Based on our experimental results over the BCI Competition IV-IIa dataset, we validated the effectiveness of the proposed framework in various aspects.