Brett Mensh, data set Ia and Ib

1a was done by averaging the DC level over the latter two-thirds or so
of the epoch for each channel (starting after the initial "notch" that
can be seen in the averaged data).  This provided one dimension for
each channel for channels one and two.  For the other channels, the
slope of the time-domain signal was used, but this turned out to be
unhelpful, so we discarded it.  Additional dimensions were supplied 
by spectral analysis using the Thomson multitaper method; channels 4 
and 6 showed differential power in the 24-37 Hz frequency range 
between the two conditions and thus were used in the final algorithm. 
All these dimensions were fed into the built-in linear classifier in 
Matlab 6 (called "classify").
Support Vector Machine did slightly better on the fully-jacknifed
training set, but was not used for the competition.

1b was done similarly, except that DC level was not used (it didn't
appear to be different between the two conditions), so it was all
spectra power.

In one of the data sets, a slight improvement was noted for
laplacian-transformed electrode data.  I'd have to look up which one
it was (and the exact frequency bands that we used), if that
information is important.

1b seemed pretty hopeless and I believe our spectral separation was
shaky at best.  I'll be amazed if anyone does better than 65% on this
data set.  But on 1a, 75% was achievable just using the DC level on
channels 1 and 2 (for the training set).