On the estimation of evoked potentials, using a feedforward neural network

We have used a multilayer perceptron to estimate the evoked potentials, masked by the EEG signal. The problem was studied on synthetic signals, generated as given in [10] and error criteria other than standard L2-norm were taken into account. We showed experimentally that, as suggested in [2], better results could be obtained this way, if the parameters were properly adjusted. An average performed on a few ensembles strongly improves the result and the number of ensembles is lower than quoted in other approaches. We have also studied the influence of the Window length and of a different number of hidden units upon the convergence speed and test error. Different activation functions having adaptive slopes were also taken into account, in order to increase the convergence speed. Though good results were obtained in this quantitative study, the trained network which resulted should be tested on real data, in order to get a complete outlook upon this problem.