[pymvpa] CRM-SPAM: [mvpa-toolbox] Re: comparing 18 classifiers on Haxby 2001 dataset
Yaroslav Halchenko
debian at onerussian.com
Tue May 25 02:00:19 UTC 2010
On Mon, 24 May 2010, MS Al-Rawi wrote:
> 1- I just wanna ask if publishing a work on the Haxby et al (2001)
> subject which is used by MVPA contains a copyright violation or not?
well... strictly speaking -- hard to say, since afaik, Dr.Haxby at that
moment has not yet released that dataset to public under any open
license (but I can be utterly wrong)
BUT, now, all subjects for that experiment are released under CC-SA
attribution v.3 and available from
http://data.pymvpa.org/datasets/haxby2001/
so, technically speaking everything is fine, attribution was in place
imho ;)
> 2- It seems that the highest accuracy value was 72.22 using 1000 voxels
> implementing random spherical SVM, and SVM weights based feature
> selection. I think this is not the highest accuracy value that one can
> get from this dataset?
do not forget that they did single trial classification... I don't
remember what I was achieving on that dataset, but 72%
single-trial for all categories (including scrambled) isn't that bad at
all: on block averages (so higher SNR, wider null-chance since smaller # of
samples) in our tutorial
http://dev.pymvpa.org/tutorial_classifiers.html#we-need-to-take-a-closer-look
we achieve 78%
--
.-.
=------------------------------ /v\ ----------------------------=
Keep in touch // \\ (yoh@|www.)onerussian.com
Yaroslav Halchenko /( )\ ICQ#: 60653192
Linux User ^^-^^ [175555]
More information about the Pkg-ExpPsy-PyMVPA
mailing list