[pymvpa] redundant voxels

[kimberg at mail.med.upenn.edu]

Yaroslav Halchenko replied:
> questions relevant to multi-variate analysis of neural data are
> always welcome to this list ;)

Thanks, I'll try not to take unfair advantage.

> do you mean smth like
> In [13]: remove_invariant_features?

As I understand it, removeInvariantFeatures just removes features that have no variance across samples.  The voxels I need to remove have different values for each sample, but neighboring voxels may have exactly the same values for each sample (these are actually 0/1 lesion maps from stroke patients, so it often happens that neighboring voxels will be lesioned in exactly the same subset of patients).  In any case, right now I do this filtering outside python for various trivial reasons.

All that said, this:

> I would expect no difference in error, iff (if and only if) you had
> "equivalent
> redundancy" among bogus and relevant features -- then it would indeed
> should not alter most of the classifiers.
[helpful example snipped]

does seem to be the right explanation.  There's obviously at least a numerical difference in the degree of redundancy between the most vs. least informative features, and it doesn't surprise me that it would vary a bit between ROIs.  To some extent, it's reassuring that the differences were so small, but in any case I think it does make more sense to prune redundant features.

Thanks for the quick reply.

dan