[pymvpa] Use searchlight algorithm only for predictions.

Roberto Guidotti robbenson18 at gmail.com
Thu Jan 31 14:53:11 UTC 2013


Thank you guys,

Yes, I know that I can't predict using a portion of voxel. Let's say that I
would like to train on full brain and test on a portion, putting out of ROI
voxel intensity to zero.
I don't know if it makes sense conceptually because I would like to predict
using a portion of features on a model built on multiple features.

Probably could be an sensitivity measure, e.g. building a classifier to
predict from flag the country for example Liberarian flag and stars and
stripes, If I use features from stripes part of the flag (common in both of
the flags) the classifier isn't able to classify - well, using a feature
selection probably those features were discarded - but using "stars" part
as ROI the classifier identifies the flag, and so I will know where the
classifier is more sensitive! (Hope I explained it clear).

I don't know if there is still a theoretical problem.

Thank you
R

PS: If I could help you to complete...

On 31 January 2013 15:38, Yaroslav Halchenko <debian at onerussian.com> wrote:

>
> On Thu, 31 Jan 2013, Michael Hanke wrote:
> >      As the subject "clearly" says, I would like to train a classifier
> using
> >      classical methods for example using all voxels and then try to
> predict
> >      using only a portion of ROI like a searchlight.
> >      I've tried to do this using the classifier as data measure in
> >      searchlight class, but obviously the features of the classifier are
> more
> >      than those used in the searchlight.
> >      How can I do?
>
> >    This is less of a technical question, but more of a conceptual one.
> You
> >    can't train an algorithm on one set of features and then run it on a
> >    different one with a different number of features.
> >    you need to have equally structured input in both training and testing
> >    stage. This could be done (think e.g. PCA projection), but whether it
> >    makes sense in you context is impossible to tell at this point.
>
> indeed!   But I guess it could be stretched to become a "technically
> legit" one in the case of kernel-based classifiers, where optimization
> and decision is done based on values within the kernel...  theoretically
> it should be possible to get the solution for one kernel (estimated on
> full data) and then apply to another (estimated on subset of the
> features)... not sure how legit it would be, but at least possible
> technically --  I guess could become an improved "sensitivity"
> measure to complement existing ones ;-)
>
> --
> Yaroslav O. Halchenko
> Postdoctoral Fellow,   Department of Psychological and Brain Sciences
> Dartmouth College, 419 Moore Hall, Hinman Box 6207, Hanover, NH 03755
> Phone: +1 (603) 646-9834                       Fax: +1 (603) 646-1419
> WWW:   http://www.linkedin.com/in/yarik
>
> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/cgi-bin/mailman/listinfo/pkg-exppsy-pymvpa
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20130131/98b63bfb/attachment.html>


More information about the Pkg-ExpPsy-PyMVPA mailing list