[Eeglablist] FastICA failure in converge

[Ronald Phlypo]

Dear Manousos,

 

 

FastICA is known to be sensitive to its initialization (the weighting matrix is randomly initialized, or specified by the user …). When a non-convergence is met, it simply starts off with a reinitialisation of the weighting vector for that component. This can easily be explained by the use of Newton-like gradient algorithms that are optimal for local minima, though not for global optimization. It is a sort of simplified Monte Carlo estimation. The best is to run fastICA multiple times and to cluster the results to obtain its consistent components (see the ICASSO package: http://www.cis.hut.fi/projects/ica/icasso/), however you will use the relatively ‘fast’ property of fastICA.

 

Personally, I think that the problem of convergence is rather due to the real data dimensionality (the number of underlying sources) than to the component’s strength (if with strength, you mean variance), since ICA is blind to source variance (it has unit variance sources as a solution, due to the ambiguity). The dimensionality of biomedical datasets is a huge topic of research, and I know that e.g. Dr. C. Hesse (Model Order Estimation for Blind Source Separation of <https://embs.papercept.net/conferences/scripts/abstract.pl?ConfID=8&Number=1763>  Multichannel Magnetoencephalogram and Electroencephalogram Signals (I)) or prof. A. Cichocki (a.o. http://www.bsp.brain.riken.jp/publications/1997/Cich-Sabala-CSCOSnolta97.pdf) have made some attempts for the estimation of the data dimensionality. 

 

Concerning the “source strength” in a different sense, note that 

1/ the strength in the objective function (the kurtosis value of the source if one uses the ‘pow3’ option) plays a crucial role and thus near Gaussian sources will be harder to recuperate than highly kurtic sources

2/ if you use the ‘tanh’ nonlinearity in fastICA, than it is only optimal for superGaussian sources; the function is thus sub-optimal for periodic triangular/sinusoidal/… or near-uniform sources! See also the extension from infomax to extended-infomax by Te-Won Lee.

 

There also exist some alternatives to the FastICA algorithm (note that it is only algorithmic, the theory remains the same), and I can advise you the RobustICA package (http://www.i3s.unice.fr/~zarzoso/robustica.html). This package has no initialization issues; it is faster, thanks to the calculation of the optimal step-size and it converges naturally with an optimal “fixed” point. Thus at least the issues of paragraph 1 in this mail could be tested.

 

 

Hope this helps,

 

 

Ronald 

 

 

From: eeglablist-bounces at sccn.ucsd.edu [mailto:eeglablist-bounces at sccn.ucsd.edu] On Behalf Of Klados Manousos
Sent: mercredi 7 janvier 2009 14:11
To: eeglablist at sccn.ucsd.edu
Subject: [Eeglablist] FastICA failure in converge

 

Dear EEGLAB users,

I use FastICA v. 2.5 in EEG signals. My problem is that fastica does not converge in 28 of my 54 datasets. When fastica stucks in one component, it tries to converge five times (making 1000 iteration each time) and finaly it appears me the following message. 

Component number ##  did not converge in 1000 iterations.
Too many failures to converge (6). Giving up.
Adding the mean back to the data.

I raised the number of interations to 10000 but then the algorithm becomes very slow (so it loses it's fast ability) and in some datasets doesn't converge again...

I would like to ask you why is that happening? 

I read that "The fastICA method failed completely to find the target sources, since the Newton iteration always converged to one of the much stronger confound sources."  and that "Even for very low signal strengths, where the fastICA method fails to converge to the target sources" <http://cds.ismrm.org/ismrm-2004/Files/000493.pdf>  here. EEG signals have very low strength...so that is the answer to my previous question or it's something more than EEG signal's strength?

Also i noticed that in the same dataset, fastica can converge ones and does not converge other time...Is it that logical?

Thank you in advnace
Have a happy new year with my best wishes

Manousos


-- 
Klados A. Manousos
Graduate Student, Research Assistant
Group of Applied Neurosciences
Lab of Medical Informatics, Medical School
Aristotle University of Thessaloniki
Thessaloniki, Greece
_________________________________________________
Tel: +30-2310-999332
Website :
http://lomiweb.med.auth.gr/gan/index_en_files/Page609.htm
Blog: http://appliedneurosciences.blogspot.com

No virus found in this incoming message.
Checked by AVG - http://www.avg.com
Version: 8.0.176 / Virus Database: 270.10.5/1883 - Release Date: 1/8/2009 6:05 PM

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20090109/bcd35361/attachment.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: Phlypo Ronald.vcf
Type: application/octet-stream
Size: 4458 bytes
Desc: not available
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20090109/bcd35361/attachment.obj>