2009 Special Issue: A linear feature space for simultaneous learning of spatio-spectral filters in BCI

摘要: It is shown how two of the most common types feature mapping used for classification single trial Electroencephalography (EEG), i.e. spatial and frequency filtering, can be equivalently performed as linear operations in space frequency-specific detector covariance tensors. Thus by first data to this space, a simple classifier directly learn optimal + filters. Significantly, if classifier's loss function convex, learning these filters convex minimisation problem. also pre-process such that resulting decision robust biases inherent EEG data. Further, based upon ideas from Max Margin Matrix Factorisation, it trace norm select solutions which have low rank. Low rank are preferred they reflect prior information about signals we expect see, classifiable contained only few spatio/spectral pairs. They easier interpret. This feature-space transformation compared with Common-Spatial-Patterns on simulated real Imagined Movement Brain Computer Interface (BCI) give state-of-the-art performance.