[pymvpa] Never-ending searchlight

[Christopher J Markiewicz]

On 01/30/2015 10:45 AM, Yaroslav Halchenko wrote:
> 
> On Fri, 30 Jan 2015, Thomas Nickson wrote:
> 
>>    Hello All,
> 
>>    I have also had this problem I think. I gave up after about a month. I
>>    though that it was just computationally prohibitive with 133 subjects but
>>    is this not the case?
> 
>>    I can run it and contribute debug data if you want as I'm still interested
>>    in the analysis.
> 
> Thank you Thomas,
> 
> Let's first see if Christopher's run might provide sufficient
> information

Naturally, the single-threaded attempt also succeeded, so I have no
useful debugging information.

My guess would have to be that there's a pathological case lurking in
the balancing, where, for some subset of balanced training sets, on a
given searchlight (probably high voxel count), the SVM fails to
converge. Since we did two balanced partitions per attempt, what would
be required was for both partitions fail to fall in the pathological
case for that searchlight.

(This is not to say that the balancer is failing but that the balancer
is why it sometimes fails and sometimes succeeds.)

-- 
Christopher J Markiewicz
Ph.D. Candidate, Quantitative Neuroscience Laboratory
Boston University

-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 473 bytes
Desc: OpenPGP digital signature
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20150130/de4be9b4/attachment.sig>