Authors:

    Michael Buschbeck <michael.buschbeck@atip.de>
    Klaus Kasper <klaus.kasper@atip.de>


Method:

    A multi-layer perceptron (MLP) comprising 144 input neurons, 2 hidden
    neurons and 3 output neurons was trained for each data set and then used to
    classify the unlabeled data.

    The input vectors were prepared by stepping through the 100-Hz data in
    intervals of 50 samples; at each step, taking 3 consecutive frames of 64
    samples of the C3, Cz and C4 electrodes; calculating the FFT for each frame;
    and reducing the resulting 32 real values to 16 input vector elements.

    The evaluation data were prepared in the same way and evaluated using the
    previously trained MLP. (Since only a single classification result can be
    calculated for each window of 50 samples, each result was replicated across
    the entire window.)