[pymvpa] Example Dataset

Per B. Sederberg persed at princeton.edu
Wed Jan 14 14:52:08 UTC 2009


On Wed, Jan 14, 2009 at 12:32 AM, Yaroslav Halchenko
<debian at onerussian.com> wrote:
>> Well, it looks like the data labels ARE shifted.  When running a
>> Bottle vs. Scissors comparison I got 50% accuracy when shifted by 3
>> and 58% correct when not shifted.
> well 50% vs 58% are within random performance margin -- I would not rely
> on them (unlike some of the well-published authors ;-))
>
Indeed ;)

>> Not convinced by that I ran a Face vs. House classification and got
>> 99.5% accuracy with no shift, but only 76.9% accuracy when shifted
>> by 3.
> that is indeed significant difference ;)
>
> btw - did you try to shift backwards?
>
No, just did the quick test of a forward shift of 3.

>> All these tests were with a Linear SVM classifier, so they should be
>> stable results.
> and default C?
>
Yup, with default C.

>> Even though all signs are pointing to the fact that they were shifted,
>> perhaps Michael knows for sure whether they were...
>
> indeed... or may be Yarik who 'recomposed' that data from the one
> provided by Haxby (or did you Michael take it from some other source)...
> and I don't really remember shifting them. In my analysis (which I've
> done long ago) I just dumped first few volumes in a block just "to be
> sure".  Dr.Haxby provided us this data some time long ago with each
> volume having a label assigned already (iirc), so it might be that they
> were already shifted by him.
>
> Sorry for not really shining the light over the whole situation ;-)
>

No prob.  I'm going to go with the idea that it's shifted b/c of the
second result, but if anyone knows for sure (nudge nudge, wink wink)
I'm sure the general populous would love to know and we could add it
to the docs.

Thanks,
Per



More information about the Pkg-ExpPsy-PyMVPA mailing list