Unsupervised Motor Imagery EEG Classification with Adversarial Mechanism

The unsupervised classification of motor-imagery electroencephalogram (MI-EEG) provides a solution to the difficulties of labeling and real scene data collecting in brain-computer interface (BCI) learning. In this study, we propose an unsupervised method based on an adversarial mechanism that transfers reliable large-scale data set information to the target domain. This approach enables us to overcome the limitations associated with the size and reliability of the data set in the target domain. The proposed model is based on a combination of feature extractor, domain discriminator, and label classifier, which allows it to produce competitive results when compared to other state-of-the-art methods. The proposed model offers the advantage of transfer ability and classification accuracy improvement. Furthermore, the inclusion of an attention mechanism strengthens the classification accuracy, yielding results that are equivalent or superior to those obtained through supervised learning. The effectiveness of the proposed architecture is demonstrated through comprehensive experiments conducted on two public datasets.