[pymvpa] searchlight concept and pyMVPA

Fri May 15 11:39:00 UTC 2009

On Fri, May 15, 2009 at 01:22:07PM +0200, Vadim Axel wrote:
> Thanks, Michael.
> Let's see if I understand correctly what you mean by "extracts boxcar-shaped
> volume". Consider a simplest case that in my fast ER design the trail length
> is one TR. You return a vector of single datapoints for each of the
> condition trials. But how you take in the account the previous event(s),
> which contaminated the mine? Do these values of your returned data points
> differ from the original values, which I used as an input or you just
> arrange them according to events?

The dataset knows nothing about the design. If you extract just one
value/volume then it won't know about the influence of preceding trials.
You have two options, which both aim to add this information. 1) you
extract more then one volume and include preceding volumes as well which
provide you with some 'reference point'. 2) You encode such information
into the labels, i.e. for a trial of condition B preceded by condition A
you use 'A->B' instead of just 'B', essentially multiplying the number
of conditions to get the number of 'transitions'.

> BTW, do you know how other software (SPM, for example) does this fast ER
> extraction? Somehow making design pseudo with optseq helps to extract the
> events.

SPM probably uses FIR (finite impulse response) functions to model the
signal change. You could do this as well, but you don't have to, since
the classifier might be able to figure it out alone. I am putting my
thesis online for you -- section 4.2.2 might give you a better idea
(although that example dataset was not fastER):

  http://apsy.gse.uni-magdeburg.de/~hanke/lit/Hanke_2009_PhDThesis.pdf

> BTW2:  the site search box doesn't trim spaces at the end of the input and
> consequently finds nothing.

I know that search engine is flaky -- that is why I've include the links
in the message.

Michael

> On Fri, May 15, 2009 at 12:39 PM, Michael Hanke <michael.hanke at gmail.com>wrote:
> 
> > Hi Vadim,
> >
> > On Fri, May 15, 2009 at 12:19:45PM +0200, Vadim Axel wrote:
> > > Hi,
> > >
> > > I feel a little bit confused with all these studies which use searchlight
> > > analysis concept, but practically make it in completely different way.
> > > Specifically:
> > >
> > > 1. Kriegeskorte et al 2006 (and Kriegeskorte et al 2007) neither talked
> > > about classification, nor SVM. His proposal was to take the output
> > (contrast
> > > image between two conditions of a standard GLM for unsmoothed data), to
> > > define a sphere, make some smart average in this sphere and then iterate
> > > over all possible spheres in order to find which spheres pass the
> > threshold
> > > (aka significant regions). No classification, no raw data.
> > >
> > > 2. Haynes et al 2007 or Soon et al. 2008 ("mind reading"), extends
> > > Kriegeskorte by taking beta vectors per condition / session and by using
> > SVM
> > > for a iterating sphere finds regions, which result in best
> > classification.
> > > No raw data and classification is based on some 10 data points per
> > > condition. The similar methodology I think was used here as well:
> > Hassabis
> > > et al. 2009
> > >
> > > 3. Some new stuff of Li at al. 2009 (
> > > http://www.cell.com/neuron/abstract/S0896-6273(09)00239-6<http://www.cell.com/neuron/abstract/S0896-6273%2809%2900239-6>
> > <http://www.cell.com/neuron/abstract/S0896-6273%2809%2900239-6>)
> > >
> > > They used Kriegeskorte searchlight method (and code as well) but then
> > they
> > > employed SVM to classify searchlight regions on averaged two volumes data
> > > points. Given that their design was fast ER I am not fully understand how
> > > this classification worked "out of the box" (I haven't succeed to dig any
> > > details from the paper).
> > >
> > > The Questions:
> > > As far as I understand, pyMVPA searchlight doesn't run univariate GLM,
> > but
> > > just runs classification for sphere in different locations?
> > > An assumption, that I have to feed it with block design data. unless I am
> > > using ERNiftiDataset, which is under development?
> >
> > Searchlight in PyMVPA is simply an abstraction of the idea of computing
> > _something_ in any sphere of a given size in the dataset. To that end it
> > doesn't matter what you want to compute. It could be a a simple GLM on
> > the mean signal, a SVM classification accuracy, or something totally
> > different -- any (in PyMVPA terminology) DatasetMeasure() should work.
> >
> > > I would appreciate if you can send any paper reference on how
> > ERNiftiDataset
> > > extracts events from fast ER design.
> >
> > Hmm, paper .... there is none (yet). Currently, the best would probably
> > be to take a look at the searchlight examples in the docs:
> >
> >  http://www.pymvpa.org/examples/searchlight_minimal.html
> >  http://www.pymvpa.org/examples/searchlight_2d.html
> >  http://www.pymvpa.org/examples/searchlight_dsm.html
> >
> > ERNiftiDataset simply extracts boxcar-shaped volume series from a 4d
> > timeseries, defined by onset and duration of events. What you do with
> > those boxcars is up to you. You could time-average them, or do more (or
> > even less) fancy things, before feeding them into a classifier.
> >
> > We are thinking about writing a manuscript that shows how to do that
> > with a really fastER design -- but that is work in progress and doesn't
> > even have the highest priority ATM.
> >
> > HTH,
> >
> > Michael
> >
> >
> > --
> > GPG key:  1024D/3144BE0F Michael Hanke
> > http://apsy.gse.uni-magdeburg.de/hanke
> > ICQ: 48230050
> >

> _______________________________________________
> Pkg-ExpPsy-PyMVPA mailing list
> Pkg-ExpPsy-PyMVPA at lists.alioth.debian.org
> http://lists.alioth.debian.org/mailman/listinfo/pkg-exppsy-pymvpa

-- 
GPG key:  1024D/3144BE0F Michael Hanke
http://apsy.gse.uni-magdeburg.de/hanke
ICQ: 48230050