An Adaptive, Sparse-feedback Eeg Classifier for Self-paced Bci

Generalized linear models (GLMs) are a very useful tool in data analysis. In this paper we project features from a dynamical system model of the EEG into a non-linear basis space. The responses from the basis functions are then mapped via a logistic classifier onto a class-posterior decision space. This mapping is parameterized via a set of weights which, importantly, we allow to be dynamically adaptive. This reflects our underlying belief that the EEG signals and subsequent decisions from BCI experiments are non-stationary. A sequential Bayesian learning paradigm gives a set of equations which may be implemented very efficiently via an extended Kalman filter (EKF). This paper shows that such adaptive classification gives good results and addresses the problem of running the method on data for which very few, or no, class labels are known such as is the case for self-paced BCI experiments.