Common spatial patterns based on generalized norms

The Common Spatial Patterns (CSP) algorithm is commonly used to finds spatial filters for classification of electroencephalogram (EEG) signals. However, conventional CSP is sensitive to outliers and artifacts because it is based on variance using L2-norm. In this paper, we consider generalized L_p norm based CSP, called CSP-L_p, and verify whichp is optimal for CSP-L_p by maximizing the Lp norm ratio of filtered dispersion of one class to the other class. The spatial filters of CSP-L_p are obtained empirically. Simulation result on a toy example shows the robustness of CSP-L_p depending on L_p-norm.