Contributors : A. Rakotomamonjy, V. Guigue Affiliation : PSI CNRS FRE-2645 INSA de Rouen Data preprocessing -------------------- * Extraction of a 667ms time-window post-stimulus for all channels. * Bandpass Filtering with a 8-order filter which cut-off frequencies are 0.1-20Hz. The filtered signal is then decimated according to the high cut-off frequency. At this point, each signal from a given channel is characterized by 14 samples. The post-stimulus signal feature vector is then composed of 14*64=896 samples. Ensemble of classifier approach -------------------------------- * Each character spelling is composed of 180 signals resulting in 180*85 training signals for a given subject. * This training signals have been partitioned in groups of 900. These 900 signals are related to the spelling of 5 characters. The original training set is then divided in 17 partitions. * A linear SVM is trained on each of the 17 partition and a channel selection procedure has been performed. The channel selection algorithm is a recursive channel elimination based on the following criterion : Criterion = TP/(TP+FP+FN) where TP,FP and FN are respectively the true positive, false positive and false negative rate evaluated on a validation set. The validation set being a subset of the 16 remaining partition sets. The C parameter has also been optimized using this validation set. Character recognition ---------------------- Given a new post-stimulus signal x which has been preprocessed, the character recognition is achieved by using as a score for a column or row intensification each SVMs output f(x) and the mixture of SVMs is achieved by summing all the scores The row and column that get the highest score is the considered as the target character after an illumination sequence. Results ------- Accuracy with respects to number of sequences Nb Sequences Subject 1 2 3 4 5 10 13 15 A 16 32 52 60 72 83 94 97 B 35 53 62 68 75 91 96 96 Mean 25.5 42.5 57.0 64.0 73.5 87.0 95.0 96.5