A co-training algorithm for EEG classification with biomimetic pattern recognition and sparse representation

In this paper, a semi-supervised co-training method is proposed to construct a high-performance classifier for motor imagery based Brain–Computer Interfaces (BCI), which is trained with little labeled data. Two novel classifiers: Biomimetic Pattern Recognition (BPR) and Sparse Representation (SR) are applied. Each classifier labels the unlabeled data for the other classifier to extend the labeled set. And the enlarged labeled set is used to generate the final classifier BPR–SR. BPR–SR is constructed by combining BPR with SR, in which SR is used to replace the traditional distance method when BPR encounters the overlap coverage problem. The features extracted by common spatial pattern were used for classification. The experiment with datasets from previous BCI competitions demonstrates the validity of our proposed co-training algorithm compared with BPR, SR and BPR–SR without co-training. And the experiment about our novel classifier BPR–SR compared with other excellent classifiers on datasets from previous BCI competitions and our own lab was also conducted. And the result shows that the classifier BPR–SR can overcome the overlap coverage problem of BPR by introducing SR.