[pymvpa] spatial normalization for MVPA - is it good or bad?

[Yaroslav Halchenko]

On Mon, 05 May 2014, J.A. Etzel wrote:

> On 5/5/2014 2:51 PM, Vadim Axel wrote:
> >I personally use always normalized and I do not think that this
> >should matter too much. I think given that normalization introduces
> >some smoothing, it may probably even increase predictions - as Hans
> >Op De Beeck showed that smoothing might be helpful for prediction
> >rate.

> Unfortunately, *should* matter doesn't always mean *does* matter, and
> I'm very hesitant to draw too many conclusions from experiences with
> smoothing: some spatial normalization algorithms are far, far
> different than Gaussian smoothing.

> That doesn't mean to never spatially normalize, but I would
> certainly never assume that it's a neutral procedure.

moreover, in case of linear "normalization" -- if you "normalize" by
applying a joined transformation (motion correction + normalization) to
each volume, instead of two separate steps (motion correction of entire
4d series and then normalization of the entire 4d at once), then those
should be nearly identical and theoretically "MVPA" metrics should be
the same if resolution is nearly the same as original and subjects
did move at least a bit since then those volumes would get
interpolated anyway to correct for the motion.

-- 
Yaroslav O. Halchenko, Ph.D.
http://neuro.debian.net http://www.pymvpa.org http://www.fail2ban.org
Research Scientist,            Psychological and Brain Sciences Dept.
Dartmouth College, 419 Moore Hall, Hinman Box 6207, Hanover, NH 03755
Phone: +1 (603) 646-9834                       Fax: +1 (603) 646-1419
WWW:   http://www.linkedin.com/in/yarik