An effective BCI speller based on semi-supervised learning

Brain-computer interfaces (BCIs) aim to provide an alternative channel for paralyzed patients to communicate with external world. Reducing the time needed for the initial calibration is one important objective in P300 based BCI research. In this paper, the training time is reduced by a semi-supervised learning approach. A model is trained by small training set first. The on-line test data with predicted labels are then added to the initial training data to extend the training data. And the model is updated online using the extended training set. The method is tested by a data set of P300 based word speller. The experimental results show that 93.4% of the training time for this data set can be reduced by the proposed method while keeping satisfactory accuracy rate. This semi-supervised learning approach is applied on-line to obtain robust and adaptive model for P300 based speller with small training set, which is believed to be very essential to improve the feasibility of the P300 based BCI.