[pymvpa] Invariant features in surface-based searchlight analysis

[c.brauchli at psychologie.uzh.ch]

Hello pyMVPA experts,

I am currently trying to set up a surface based searchlight analyses as documented
in the pymvpa manual (http://www.pymvpa.org/examples/searchlight_surf.html), but on
structural data. I am trying to classify two groups (A,B) with ten subjects each. 
My input data contains values for "local gyrification" in voxels that lie between the white matter (%s.smoothwm.asc) and pial (%s.pial.asc) surfaces. All other voxels are zero. 
I masked my input data with my query engines voxel selection ("qe.voxsel.get_mask()") which gives me a final datastructure of (20, 235866). Yet, I have many invariant zero-value voxels, e.g. in regions of the corpus callosum where white and pial surfaces collapse on each other and thus no values for local gyrification are assigned. I tried removing the invariant features with "remove_invariant_features", but this gives me an error when running the searchlight analyses as the new datastructure (20, 206197) does not fit the voxel selection of the queryengine.
When I run my searchlight (using LinearCSVMC classifier), I get the following warning for some voxels:

#    [SLC] DBG:                   +0:00:08 _______[0%]_______ -1:41:17  ROI 38 (38/27307), 100 #    featuresWARNING: Obtained degenerate data with zero norm for training of <LinearCSVMC>.  #    Scaling of C cannot be done.

I guess this comes from the invariant voxels in my dataset, yet I see no possibility to exclude them from my analyses.

Also, my final accuracies are centered around 0.1 and not as expected around the chance level of 0.5.

Are there other classifiers that deal better with invariant features or should I rather run an "old-fashioned" spherical searchlight analyses where I could use "remove_invariant_features"?

Thanks in advance for your help!
Christian 

-- 
Universität Zürich
Christian Brauchli, MSc 
Psychologisches Institut
Neuropsychologie 
Binzmühlestrasse 14, Box 25 
CH-8050 Zürich

+41 44 635 74 51 Telefon
+41 44 635 74 09 Telefax

www.psychologie.uzh.ch
c.brauchli at psychologie.uzh.ch
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20170901/dfe92a07/attachment.html>