Bayesian common spatial patterns with Pitman-Yor process priors

Common spatial patterns (CSP) and probabilistic CSP (PCSP) are popular methods for extracting discriminative features from electroencephalography (EEG), but they are trained on a subject-by-subject basis so that inter-subject information is neglected. When only a few training samples are available for each subject, the performance is degraded. In this paper we present a method for Bayesian CSP with Pitman-Yor process (PYP) priors, in which spatial patterns (corresponding to basis vectors) are simultaneously learned and clustered across subjects using variational inference, allowing for a flexible mixture model where the number of components are also learned. Spatial patterns in the same cluster share the hyperparameters of their prior distributions, so that the information transfer is encouraged between subjects involving similar spatial patterns. Numerical experiments on BCI competition IV 2a dataset demonstrate the high performance of our method, compared to existing PCSP and Bayesian CSP with a single prior distribution.