[pymvpa] SVM classification of data with temporal correlation

[Francisco Pereira]

Vadim,

I don't think it would hurt to use correlated examples for *training*
the classifier. In my experience averaging a few consecutive TRs seems
to be advantageous, both for performance (it should reduce noise) and
for lowering training time. For testing, though, if you want to do a
binomial test of the accuracy (under the H0: classifier performs at
chance) you need to have independent examples (if not, the <n> in that
test will effectively be smaller than the number of examples and your
p-value lower than it would otherwise be). One example per block seems
reasonable, if the blocks are well separated (let's say 8-10 sec). If
you want to be extra-careful, make sure you never use examples for
training that are too close in time to examples used for testing. One
way of doing this is to do cross-validation leaving one run out as
test at each time.

cheers,
Francisco