[pymvpa] ERNiftiDataset with different event durations

[Yaroslav Halchenko]

> OK, but my feeling was, that the classifier has more problems to
> focus on the differences between conditions A and B if other
> (irrelevant) features are included in the pattern vector. To test
Well... they are irrelevant if they are irrelevant, but here I would not
name the presentation  of the target as irrelevant factor, if it was not
counter-balanced. Good that it was ;)

> this, I performed the analysis only with examples covering the 6sec.
> Delay, so the Boxcarmapper included only relevant examples of class
> A and B. Than I changed the event duration just for one second with

>     for ev in evs:
>        ev['duration'] = 7

> and the accuracy substantially decreased (may be this effect could
> be "undone" with correlation stability analysis before
> classification).
So, you did binary classification, right?

was it full brain or some ROI? what # of voxels?

TR was 2 or 3 seconds so changing duration from 6 to 7 added another
sample into 'spatio-temporal' pattern, right? Depending on the
classifier, the effect of substantial increase of number of feature
could have confused it, unless you've used some classifier with feature
selection... so what classifier was it (family/parameters/etc)

-- 
                                  .-.
=------------------------------   /v\  ----------------------------=
Keep in touch                    // \\     (yoh@|www.)onerussian.com
Yaroslav Halchenko              /(   )\               ICQ#: 60653192
                   Linux User    ^^-^^    [175555]