Unsupervised MEG classification by Riemannian Geometry and class centroid matching

In brain-computer interface, information decoded from brain can be used as control signal. However, classifier trained from previous subjects suffer from performance drop due to some factors, including environmental, physiological and instrumental changes. Concretely speaking, statical distribution alters across subjects, as well as sessions. In this paper, we focus on across-subject problem. Thus, we propose a framework for decoding MEG and matching data distribution across subjects. First, features from Riemannian tangent space were extracted to build common feature space between source domain and target domain. Second, taking advantages of clustering in target domain, supervised subspace learning in source domain and matching the class centroid between two domains, we proposed an improved transfer learning method named class centroid matching (CCM). Several experiments had been conducted on a MEG dataset, which shows that our proposed method is effective to reduce discrepancy across subjects and can achieve a promising performance than other comparable methods.