A multi-class brain-computer interface using large margin nearest neighbor classification

Motor imagery based brain-computer interface (BCI) translates subject´s motor intention into a control signal through electroencephalogram (EEG) pattern classification. In this paper, a large margin nearest neighbor (LMNN) method is applied for the classification of multi-class BCI based on motor imagery. The main idea of LMNN is to learn a Mahalanobis distance that tries to collapse examples in the same class to a single point, meanwhile keeps examples from different classes far away. Here, we present a modification to LMNN method so that the computational complexity is significantly reduced. Experimental results on Data Set 2a of BCI Competition 2008 show good performance of the method. Besides high classification accuracy, LMNN method also has the advantage of requiring no modification or extension for multi-class classification from binary case.