[pymvpa] search-light vs. ROI analysis: significance puzzle

Malin Björnsdotter malin.bjornsdotter at neuro.gu.se
Tue Feb 26 12:48:44 UTC 2013


Hi Vadim!

I sequentially added more and more search volumes (lights),
approaching the number in the exhaustive map. At some point
(surprisingly early) in this process this random average map was
better (= had a larger area under the receiver operating
characteristic curve) than the non-average, exhaustive (searchlight)
map on simulated data. I have attached a hotted up version of a figure
from the Neuroimage paper, that may make it more clear: at less than
5,000 classifiers (=searchlights) the Monte Carlo approach performed
at the same mapping level as the exhaustive (searchlight) algorithm
(which required over 25,000 classifiers, i.e. as many voxels as there
were in the brain). So, the number of search volumes you choose
depends on your definition of satisfactory performance. :)

Also, there is a trick - in my approach, the search volume selection
is not quite random. I made sure that every voxel was included in the
same number of search volumes, i.e. I partitioned the entire brain
into search volumes (some much smaller than that specified by the
radius parameter) such that every voxel was included in one.

~Malin

On Tue, Feb 26, 2013 at 8:09 PM, Vadim Axel <axel.vadim at gmail.com> wrote:
> Indeed, very similar - I only make it not random, but rather sequentially
> iterating over all brain. In such a way each voxels participates in roughly
> the same amount of lights. I could not figure out from paper, Malin, in how
> many lights the voxel should participate in order to achieve satisfactory
> performance?
>
> On Tue, Feb 26, 2013 at 3:47 AM, Malin Björnsdotter
> <malin.bjornsdotter at neuro.gu.se> wrote:
>>
>> > In my method, the hitrate is assigned to all the
>> > voxels in the light and given that each voxel participates in many
>> > lights,
>> > the hitrates are averaged. So, using my method a voxel hitrate reflects
>> > many
>> > possible environments. I try to compare the results of both and so far
>> > it
>> > seems that with your method the results are more patchy.
>>
>> That sounds pretty much exactly as what I've been doing. :-) Jo has an
>> excellent blog entry about this:
>> http://mvpa.blogspot.sg/2012/09/random-searchlight-averaging.html
>>
>> _______________________________________________
>> 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 --------------
A non-text attachment was scrubbed...
Name: Fig3.png
Type: image/png
Size: 19346 bytes
Desc: not available
URL: <http://lists.alioth.debian.org/pipermail/pkg-exppsy-pymvpa/attachments/20130226/82f40ffd/attachment-0001.png>


More information about the Pkg-ExpPsy-PyMVPA mailing list