Jeremy Hill (Max Planck Institute for Biological Cybernetics, Tuebingen) and
Michael Schroeder (Computer Science department, Tuebingen University).
We interpolated each nan-sample linearly, then downsampled the data to
100 Hz with a smooth falloff from 45 to 50 Hz, then did linear
detrending, then performed infomax ICA, then obtained amplitude spectra
using Welch's method with 5 overlapping hanning-windowed segments, then
performed linear PCA on all the resulting features. We then split the
problem up into binary contrasts and trained linear SVMs. For each
binary classification, we performed 10-fold CV to find the best
regularizer and the optimal number of PCs (PCs were ordered by a
ROC-based method). We performed all pairwise contrasts, all
one-against-rests, and all contrasts between merged pairs of classes,
for a total of 13 output values. These were then decoded to 4 class
scores using loss-based decoding. The scores are constant for the
entire duration of each trial.