[pymvpa] Noise normalization

[yoh at onerussian.com]

On Wed, 01 Apr 2020, Raúl Hernández wrote:

>    Dear PyMVPA community,
>    I'm a bit confused with noise normalization and I was hoping someone can
>    clarify it for me. Walther et al., (2016) mentions that noise
>    normalization is an important step in multivariate fMRI. For what I
>    understand, if I Z-score each voxel, I'm performing a univariate noise
>    normalization, as I'm transforming the signal and it's variations to a
>    score that accounts for the dispersion of the data (thus, the noise). Is
>    this correct? Or the noise normalization is something different?

Hi Raúl,

zscoring could indeed be considered to be a "poor man" univariate
noise normalization if you zscore against some condition of no interest
(e.g. rest), where mean/var would be estimated not across the entire
time series, but only where there is presumably no signal of interest
(thus only noise).

If you take it across the entire time series, then it would indeed be
hardly "noise normalization", but just "data standardization" (bring it
to the common mean of 0 and variance of 1).

In that paper (https://doi.org/10.1016/j.neuroimage.2015.12.012) noise
normalization seems to not involve mean subtraction (like zscoring does)
and then in a univariate case would just divide by the noise variance.
In multivariate case -- would use entire covariance structure, thus
would decorellate  (not necessarily "approximately independent" as
authors say) remaining noise across voxels.

Hope this helps

-- 
Yaroslav O. Halchenko
Center for Open Neuroscience     http://centerforopenneuroscience.org
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