[pymvpa] Sensitivity analysis with GNB

[Yaroslav Halchenko]

On Fri, 12 Dec 2014, Thomas Nickson wrote:

>    I've trained a GNB classifier on a dataset with moderate success and would
>    like to look at the regions that the classifier finds most interesting.

just use sphere_gnbsearchlight  ? ;)

> I
>    notice in the code that there is a sensitivity analyser for the GNB module
>    but that it has been removed:

;) wonders of the open-source, aren't they? ;)

>    ## class GNBWeights(Sensitivity):                                        
>    ## """`SensitivityAnalyzer` that reports the weights GNB trained         
>    ## on a given `Dataset`.                                                 
>    ## """                                                                   
>    ## _LEGAL_CLFS = [ GNB ]                                                 
>    ## def _call(self, dataset=None):                                        
>    ## """Extract weights from GNB classifier.                               
>    ## GNB always has weights available, so nothing has to be computed here. 
>    ## """                                                                   
>    ## clf = self.clf                                                        
>    ## means = clf.means                                                     
>    ## XXX we can do something better ;)                                     
>    ## return mean                                                           

>    Is the use of the means considered to be poor in some sense?

well -- kinda since they wouldn't be anyhow describing the 'sensitivity'
really, just a mean of the voxel given a class  label.

>    Could anyone provide more information about this:

>    # XXX Later come up with some                       
>    # could be a simple t-test maps using distributions 
>    # per each class                                    

yeah -- something like that might prove being useful, but probably not
too far from omnibus or ad-hoc Anovas which we have.  So that is why
they were never implemented

-- 
Yaroslav O. Halchenko, Ph.D.
http://neuro.debian.net http://www.pymvpa.org http://www.fail2ban.org
Research Scientist,            Psychological and Brain Sciences Dept.
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