Albany dataset (IIb, P300) D.M.J.Tax and B.Blankertz Fraunhofer FIRST.IDA Kekulestr. 7, D-12489 Berlin, Germany davidt@first.fraunhofer.de Problem A user is presented a 6x6 matrix of characters. Randomly each row and column is highlighted 15 times. When the user focuses on one letters, a highlighting of the correct row or column evokes a P300 response. The task is to find the letter the user was focusing on from the user EEG responses of the 15 repetitions of the 12 highlights. Preprocessing From the original data, the channels F3, F1, Fz, F2, F4, FC3, FC1, FCz, FC2, FC4,C5, C3,C1,Cz,C2,C4,C6,CP3,CP1,CPz,CP2,CP4, P7, P8,AFz were taken (25 channels in total), and epochs around the row/column highlights were extracted (from -50 to 500 ms). A baseline-correction was made by subtracting the average signal taken between 0 and 150 ms. Then the signal between 250 and 451 ms. was 10 times subsampled, and 5 values per channel were obtained. This 5×25-dimensional feature vector was used in the classifier. In total 7560 objects were present in the training set (that means 42 letters). Classifier and combining The best classifier for the individual P300 responses is the regularized linear discriminant, with a regularization constant of 0.0375. This was obtained using 10 times 10-fold crossvalidation. The 12x15 epochs for each letter were classified. Each repetition of 12 epochs was resorted to get the 6 row and 6 column response in standard order. Thus a matrix x of 12x15 outcomes were obtained. The lowest and highest 3 classifier outcomes were ignored, and from the rest the average was taken: sortx = sort(x,2); % sort the 15 classifier outputs x_ave = mean(sortx(:,4:12),2); % remove the extrema and average Finally, the maximum for the 6 rows and the maximum for the 6 columns was taken as the final output. These indices were used to find the letters in the letter-matrix. The results The results (using 5-fold cross-validation) on the training set was 3 errors out of 42: C G* T D O G F I S H W A T E R B O W L H A T G* A T G L O V E S H O E Y* F I S H R A T (the letters indicated by * are errors, the classifier outputs do not show much hope for improvement here). The outcome of the test set is F O O D M O O T H A M P I E C A K E T U N A Z Y G O T 4 5 6 7