[pymvpa] Example Dataset

[Yaroslav Halchenko]

since princeton's MVPA people also distribute haxby data -- I looked at
their labels -- there is only 5 baseline volumes at the beginning, so I
guess I was wrong at reincarnating the ones on both sides, and should
have reincarnated 2 on front instead to provide "not shifted" labels.

so, effectively we have labels shifted by 1 volume (TR=2.5). taking into
account that some slices are acquired at the end of TR duration, we do
have relative shift from 2.5 to 5.0 (depending on the slice) seconds
relative to the onset of the stimuli. That is imho sufficient for
response to reach 'significant' level -- and to acquire good
generalization for basic classes (e.g house-vs-face) as Per got.

indeed, should be mentioned in the docs somewhere ;)

On Wed, 14 Jan 2009, Per B. Sederberg wrote:

> Thanks Yarik!

> That does get us closer, but it doesn't explain why I get near perfect
> accuracy with the non-shifted labels for face vs. house (I guess it
> could just be guessing correctly on the other labels, which are
> probably rest with residual face or house activity.)

> I'll see how well I can do with shifts of 2 and 1 (to line up with
> shifts of 3 and 2 when you take into account the dropped label that
> you just uncovered.)

> Best,
> Per



> At Wed, 14 Jan 2009 19:21:26 -0500,
> Yaroslav Halchenko wrote:

> > ok -- to shine the light a bit.

> > In our paper's repository (don't you have it, Per? ;-)) I found 
> > tools/haxby_data_convert.py

> > which I've used to reconstruct proper 4D Nifti file and labels file out
> > of the data we got from Haxby awhile ago. In original dataset from Haxby
> > we had 3D volumes in separate files and text files listing volumes
> > filenames per each category (except baseline condition). From reading
> > the original paper, and comparing to the labels we got, it looked like
> > Haxby discarded 1 volume on each side of the block of stimuli. That
> > script tried to reincarnate those labels by inserting proper label at
> > the beginning and at the end of the block.

> > So, theoretically, those labels we have in example dataset ARE NOT
> > shifted accordingly by HRF, unless my logic in reincarnating the labels
> > was wrong, that Haxby removed just volumes from the beginning of the
> > block, then what we have now would account for HRF by a shift of 1
> > volume.

> > I guess this would be the further I could dig out the history of the
> > labels ;-)

> > -- 
> > Yaroslav Halchenko
> > Research Assistant, Psychology Department, Rutgers-Newark
> > Student  Ph.D. @ CS Dept. NJIT
> > Office: (973) 353-1412 | FWD: 82823 | Fax: (973) 353-1171
> >         101 Warren Str, Smith Hall, Rm 4-105, Newark NJ 07102
> > WWW:     http://www.linkedin.com/in/yarik        


-- 
Yaroslav Halchenko
Research Assistant, Psychology Department, Rutgers-Newark
Student  Ph.D. @ CS Dept. NJIT
Office: (973) 353-1412 | FWD: 82823 | Fax: (973) 353-1171
        101 Warren Str, Smith Hall, Rm 4-105, Newark NJ 07102
WWW:     http://www.linkedin.com/in/yarik