Multi-Source Fusion Domain Adaptation Using Resting-State Knowledge for Motor Imagery Classification Tasks

Abstract

Transfer learning is the method that makes use of knowledge from other fields to solve problems in related fields. It has been shown that it can deal with the problem of insufficient labeled data for new users or new tasks in the brain-computer interface. Domain adaptation is one of the transfer learning methods which is widely used for its excellent performance. Here, the offline cross-subject EEG signal classification is mainly focused on. The unlabeled EEG trials of the new user are classified by using the EEG trials with labels from source subjects. In this paper, a novel transfer learning method called multi-source fusion adaptation regularization (MFAR) is proposed. MFAR preprocesses the EEG signal by aligning the motor imagery trials to their resting state trials, and can reduce the differences among subjects. It also defines a learning framework by combining weighted balanced distribution adaptation (W-BDA), source empirical risk, and manifold regularization to further reduce the variation between source and target domains. We validated the method on two BCI Competition IV datasets for motor imagery tasks. In the absence of labeled EEG trials of the target subject, compared with the excellent counterparts, the classification accuracy increases by 9.28% and 11.73%. After the alignment algorithm is added, the accuracy of the MFAR is improved by 9.36% and 4.17% on the basis. The experimental results show that our learning framework outperformed several state-of-the-art transfer learning algorithms. Even when the training data from the new user are sufficient, the proposed approach achieves good performance.