Normalize data by subtracting the mean across voxels
per chunk per condition (target).
Split data into a training set (about 75% of all values) and a testing
set (about 25% of values), unless there are only two runs, in
which case it is 50% training and 50% testing.
For each pair of conditions, train the classifier.
Then test on the average of the testing set, i.e., only on two
samples. This trick usually boosts the performance (credit:
Hans P. Op de Beeck)
Args:
evds (event-related mvpa dataset)
Kwargs:
nIter (int, default: 100)
Number of random splits into a training and testing sets.