[pymvpa] Hyperalignment: SVD did not converge

Kiefer Katovich kieferk at stanford.edu
Tue May 1 00:13:01 UTC 2012

An update to my last question:

I decided to do an iterative test of hyperalignment to see when SVD
would not converge and when it would. In total I am now testing on 21

I randomized the ordering of the subject, started with the first two
in the list and attempted to hyperalign them together. If
hyperalignment succeeded, I kept that subject in the list and added
the next subject in. If a subject failed, then I would remove him from
the batch of subjects to hyperalign.

Out of the 21 subjects, only a group of 6 subjects managed to
successfully hyperalign to each other by the end. The other subjects
would fail due to SVD not converging. I will have to run this again to
see if the same 6 subjects end up in the hyperalignment group or if
there is a different group this time.

Regardless, it seems that the SVD not converging happens very often
with this data. The data is not very obscure: subjects decide between
two different gambles on the screen and then see the outcome of the
gamble. The TRs are classified by trial type, and I am only looking at
the first 2 TRs of each trial, prior to their decision.

Could the SVD non-convergence be a classification error on my part?
I'm not sure.

>From your previous response in a different thread you mentioned that
SVD does not converge when it is unable to find an appropriate dataset
to kick off hyperalignment. This doesn't quite make sense to me since
some new subjects added to the pool cause SVD non-convergence, while
adding others allows hyperalignment to finish. Shouldn't it be able to
choose one of the older subjects that was already successfully run
through hyperalignment? Does it have to find an initial subject that
somehow "matches" all the subjects in the pool?

Sorry, I'm not too clear on what SVD requires from the datasets to converge.


More information about the Pkg-ExpPsy-PyMVPA mailing list