Preprocessing: A 6th order Butterworth filter was used with a passband of 0.1-40Hz. The 15 subtrials were ensemble averaged in groups of 3 (5 subtrials each), to reduce entropy. In the second case, the first 5 subtrials were used as is, without averaging. A 600 ms window was picked 100 ms after stimulus i.e., thw indow was 100-700ms. The 144 samples in this window were further decimated(subsampled) by a factor 0f 16. Thus there are 144/16 = 9 features per channel. Finally, each training vector is unit normalized. Channel Selection by Fisher criterion. The Fisher score for each feature was computed and averaged by channel. The channels were then ranked by the averaged channel scores. The best 20 channels for each subject were used. Classification by Kernel Fisher discriminants using a Gaussian kernel. Each of the subtrials per stimulus was projected onto the discriminant and the scores were averaged. Rows and columns with maximum averaged scores were selected as targets.