[pymvpa] searchlight

[Christopher J Markiewicz]

On 05/15/2015 09:25 AM, basile pinsard wrote:
> Hi MVPA experts,
> 
> I have a theoretical question that arised from recent analysis using
> searchlight (either surface or voxel based):
> 
> What is the most sensible feature selection strategy between:
> - a radius with variable number of features included, which will make
> the different classifiers trained on different amount of dimensions;
> - a fixed number of closest voxels/surface_nodes that would represent
> different surface/volume/spatial_extent depending on the localization.

From my reading, the more common is the former. This is probably
because, without evidence that your results particularly correlate with
searchlight size, the more interpretable figure is one in which each
voxel represents a statistic taken over a fixed spatial extent.

A fixed number of voxels (I agree with Jo that one should always use
voxels; even if you are using surface nodes to define a neighborhood,
these should be mapped back to voxels to avoid smoothing and resampling)
is beneficial if you are using an error metric that is sensitive to
dimensionality, such as mean squared error.

With a surface searchlight of radius 9mm, I get a distribution of
searchlight sizes (in one subject) that's approximately normal(66, 8). I
have not found that the cross-validation training error of classifiers
(linear SVM, mostly) is particularly sensitive to searchlight size. On
the other hand, attempting to use the same searchlights with regression
problems produces results that correlate strongly (positively or
negatively, depending on regression algorithm) with number of voxels.

> I had the examples with surfaces, for which I used a spherical templates
> (similar to 32k surfaces in HCP dataset) transformed into subject space.
> I computed the number of neighbors for each node with a fixed radius and
> noted a differential sampling resolution in the brain, which somewhat
> overlay with my network of interest (motor) and thus my concerns.

Do your preliminary results correlate with searchlight size across
several regions? That would be my primary indication that this is a concern.

> With voxel based searchlight, depending on masking voxels on the borders
> of the mask will have less neighbors in a fixed radius sphere.

Could you smear the mask with your searchlight, i.e. extend it in all
directions? You'll still be including (presumably) uninformative voxels,
but at least you won't be dimensionality itself that gets you.

> PyMVPA has only this strategy for now, but I read many papers with fixed
> amount of features in Searchlight.
> 
> What do you think?
> 
> I did an ugly modification to have a temporary fixed feature number
> (closest) on surface but it should be optimized:
> https://github.com/bpinsard/PyMVPA/commit/1af58ea8a57882ed57059491c19d83bed43e0bce

If I'm reading this right (I haven't dug into the PyMVPA surface
searchlight implementation), this is selecting a maximum number of
surface nodes, and then mapping that to voxels, and will still end up
with variable numbers of voxels, depending on the density of surface nodes.

What about sorting nodes based on distance, mapping to voxels, and then
taking the first max_features voxels?

-- 
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: 455 bytes
Desc: OpenPGP digital signature
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20150515/714e1a36/attachment.sig>