Toward Cross-Brain-Computer Interface: A Prototype-Supervised Adversarial Transfer Learning Approach With Multiple Sources

Jia Zhang, Jinglong Fang, Siwei Liu, Dezheng Liu, Hanrui Wu, Jinyi Long

Abstract: Transfer learning is useful in increasing the generalization ability of a model, for dealing with variations among different subjects in the brain-computer interface (BCI). Nevertheless, most of the existing methods are either implemented only considering the individual differences or half-baked in the intrasubject nonstationarity exploitation. How to generate more effective models for cross-subject classification in BCI by considering both the intrasubject and intersubject variances based on the weak and nonstationary electroencephalogram (EEG) signals is still difficult. In addition, it is a high-cost process to acquire a mass of subject-specific labeled EEG data for modeling. To cope with these challenges, we present a novel prototype-supervised adversarial transfer learning approach with multiple sources, named PSAT. Specifically, based on each source domain, PSAT involves a domain discriminator to reduce the intersubject discrepancy and a prototype mapper to constrain the intrasubject nonstationarity. Consequently, we obtain the domain invariant representation to make the prediction. Furthermore, a unified framework is designed for optimizing EEG classification performance through a weighted combination of predictions from multiple source domains. This multisource optimization approach enables the effective integration of diverse information sources, leading to improved prediction accuracy. We compare PSAT with several other well-established EEG classification methods on three EEG-based motor imagery competition datasets. The proposal can perform the best on all the datasets. These results show the effectiveness of PSAT in improving classification performance by learning the intersubject discrepancy and the intrasubject nonstationarity simultaneously, which also can fully utilize multisource data to achieve the purpose. In a nutshell, this study marks a crucial step toward the practical implementation of EEG classification in real applications.