Please note that we do not use ICA for this data set but the basic
method based on Tikhonov Regularization can be found here:
D. K. Wong, M. Perreau-Guimaraes, E. T. Uy, and P. Suppes,
"Classification of individual trials based on the best independent
component of EEG-recorded sentences," Neurocomputing, vol. 61, pp.
479-484, Oct 2004.
In terms of preprocessing, each trial of the data was downsampled and
rescaled to have a range between -1 to +1. A portion of the trials
(70%) were used for "training" and the rest of the trials were used
for validation. "Training" is done by computing a linear model with
Tikhonov regularization: w = (X'X+lambda2*I)^-2*X'y where y is a
vector of the target values (1 denoting the class corresponding to w
and 0 otherwise), with yi corresponding to the trial xi, where xi is a
row vector of matrix X. The parameter lambda is the regularization
parameter and w is the weight vector corresponding to the class.
Classification, in general for K-class problem, is computed by
choosing the class with the maximum inner product for a test
trial xt. With the knowledge of the true labels, classification rates
were computed simply by dividing the number of correct matches over
the sum of all the test trials.
Classification rates of the validation trials were computed using X
based on each channel and the channels were sorted accordingly.
turn determine the size of X) and the value of lambda were optimized
in a similar manner.
By estimating lambda and defining data matrix X with the selected
channels(30 channels were used for this particular submission), the
linear model was recomputed using all the labelled data (all the
training and validation trials). The weights were then applied to the
test trials for estimating all the labels of the test trial.