[Eeglablist] paradox in whitening

[Makoto Miyakoshi]

Dear ICA persons and list,

Mo observed FastICA showed more mutual information reduction than Infomax.
I'm curious to know if FastICA outperforms Infomax under some conditions.
Any comment?

Makoto

2013/3/19 Mo Khalili <mokhal at yahoo.co.uk>

> Dear Makato,
>
> I am using the freely available dataset of the same paper (Delorme et
> al.(2012) in PLoS One). Full details of the experimental paradigm and
> dataset is provided in the following link:
> http://sccn.ucsd.edu/wiki/BSS_Comparison.
> Actually I only calculated the averaged absolute correlation between
> estimated sources.
>
> Thanks
> Mo
>
>   ------------------------------
> *From:* Makoto Miyakoshi <mmiyakoshi at ucsd.edu>
> *To:* Mo Khalili <mokhal at yahoo.co.uk>
> *Cc:* EEGLAB List <eeglablist at sccn.ucsd.edu>
> *Sent:* Tuesday, 19 March 2013, 16:51
>
> *Subject:* Re: [Eeglablist] paradox in whitening
>
> Dear Mo,
>
> Ok, why don't you tell us about your data (number of channels, data
> length, task, recording system, etc). We may be able to find a reason.
>
> Makoto
>
> 2013/3/18 Mo Khalili <mokhal at yahoo.co.uk>
>
> Hi Makoto,
>
> yes, I used the same dataset with multiple runs of ICA (different
> decompositions); there wasn't any significant change in the results.
> the absolute correlation between pairs of components ,resulting from
> FastICA, is near zero, so the components are well decorrelated. However, in
> the infomax the result shows a significantly higher correlation between IC
> pairs.
>
> Mo
>   ------------------------------
> *From:* Makoto Miyakoshi <mmiyakoshi at ucsd.edu>
> *To:* Mo Khalili <mokhal at yahoo.co.uk>
> *Cc:* "eeglablist at sccn.ucsd.edu" <eeglablist at sccn.ucsd.edu>
> *Sent:* Tuesday, 19 March 2013, 0:23
> *Subject:* Re: [Eeglablist] paradox in whitening
>
> Dear Mo,
>
> That's not consistent with Delorme et al. (2012) in PLoS One. I wonder why
> too. Did you use the same data to test two different weight matrices?
>
> Makoto
>
>
>
> 2013/3/18 Mo Khalili <mokhal at yahoo.co.uk>
>
> Hi,
>
> I have a general query about performance of ICA algorithms.
> As I know FastICA is trying to estimate sources based on maximizing the
> non-gaussianity while, infoamx is trying to estimate the components based
> on minimizing mutual information between sources. What wonders me is , when
> I calculated the mutual information between ICs, the averaged mutual
> information between pairs of ICs in  FastICA is lower than ICs estimated by
> infomax. Additionally when I calculate the correlation between pairs of the
> estimated ICs, the averaged absolute correlation between ICs obtained by
> FastICA is significantly lower than infomax.
> in my opinion this result speculates, the whitening(sphering) step in the
> infomax is different from FastICA, and it does not perform the same.
> However, when I look at the codes of FastICA and infomax, both have used
> the same method (and functions) for whitening.
> I appreciate it if anyone can help me through understanding the reason.
>
> Regards,
> Mo
>
>
>
>
>
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to
> eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to
> eeglablist-request at sccn.ucsd.edu
>
>
>
>
> --
> Makoto Miyakoshi
> JSPS Postdoctral Fellow for Research Abroad
> Swartz Center for Computational Neuroscience
> Institute for Neural Computation, University of California San Diego
>
>
>
>
>
> --
> Makoto Miyakoshi
> JSPS Postdoctral Fellow for Research Abroad
> Swartz Center for Computational Neuroscience
> Institute for Neural Computation, University of California San Diego
>
>
>


-- 
Makoto Miyakoshi
JSPS Postdoctral Fellow for Research Abroad
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20130322/71634821/attachment.html>