[Eeglablist] RegIca AR and SCRLS algorithm

Wed Apr 6 07:51:21 PDT 2011

Hi Klados,

yep, the problem is running regression on components. It seems to me that
for some reasons it tends to be generally more unstable than performing same
algorithms on the source signal (chans x samps). Could be for the presumably
bigger variance in magnitudes that we have between comonent and comoponent
with respect to the one between chans and chans?

And when one forces stability [SCRLS] on about 45 minutes [reshaped signal
of interest from 3D (chans x samps x trials), 2d (chans x [samps,trials])]
 of 512Hz 69 chans EEG, using SCRLS_regression.m, it is practically
impossible to keep numerical precision at 50 bits.

I had to go down to 30-34 bits (currently trying to do that at 30, which is
sensibly faster than 32 and 34), and I was wondering how this trade off in
numerical precision will in effect affect the quality of the regression, and
if and how playing with sigma and labmda may produce better results.

Thank you for your skype contact, I am going to run some more tests and I
will contact you with more impressions.

Mahesh

ps> Of course I was meaning Klados, I think I got confused by some previous
messages.

Mahesh M. Casiraghi
PhD candidate - Cognitive Sciences
Roberto Dell'Acqua Lab, University of Padova
Pierre Jolicoeur Lab, Univesité de Montréal
mahesh.casiraghi at umontreal.ca

I have the conviction that when Physiology will be far enough advanced, the
poet, the philosopher, and the physiologist will all understand each other.
Claude Bernard

2011/4/5 Μανούσος Κλάδος <mklados at gmail.com>

> Dear Mahesh,
>
> Let me understand something. In this kind of dataset ICA is running and the
> problem is with the regression part? If
> so you can perform reg in smaller segments of ics... Also you are able to
> try another regression scheme based one a single step regression and not in
> adaptive filters...
>
> Also watch your eog signals ... You have 4 channels but from these channels
> 2 bipolar signals are obtained  which are going to be used as an input in
> every adaptive filter...
>
> Everything you need About regica I am available for all of you in my e-mail
> or even better at Skype(mklados) you can also follow me at twitter
> (@mklados) where future versions of regica will be announced as well as many
> other artifact rejection and neuroscientific staff will be pointed
>
> Sincerely yours
> Manousos Klados
>
> P.S.
>
> Which is Florian?
>
>
> ___________________________
> Manousos Klados
> PhD Candidate
> Group of Applied Neuroscience
> Lab of Medical Informatics
> Medical School
> Aristotle University of Thessaloniki
> ___________________________
> iPhone
>
> 4 Απρ 2011, 11:10 μ.μ., ο/η Mahesh Casiraghi <mahesh.casiraghi at gmail.com>
> έγραψε:
>
> Dear EEGLabbers,
>
>
> I am trying here to test if the hybrid methodology proposed by Florian and
> his group [<http://lomiweb.med.auth.gr/gan/mklados/index.php?option=com_k2&view=item&id=25:regica>
> http://lomiweb.med.auth.gr/gan/mklados/index.php?option=com_k2&view=item&id=25:regica]
> may be effective in removing eye-artifacts in an experiment where 10 secs
> epochs need to be segmented.
>
>
> The concept of the methodology seems promising to me, but I am nonetheless
> a bit puzzled with respect to which regression algorithm might be adopted.
> When it comes to run the code on a real subject (512Hz sampling rate,
> continuous data, 64 EEG and 4 EOG chans, about 1.30 hours) LMS and CRLS
> become both unstable and fail after just few steps. H INF ew and tv
> algorithms, do that too. It seems to me the only option left is to make use
> of SCRLS_regression.m from the AAR toolbox, but as the relative
> documentation suggests, the function is not really optimized for fast
> computation, and the reg procedure seems to take ages to converge [running
> it in matlab 64, on a 4cores pc, for a 1.30 hours continuous EEG and it is
> still trying to converge after 23 hours of computation, just one processor
> used].
>
>
> Question is: can someone out there with some experience with AAR toolbox
> and/or SCRLS algorithm provide some insights on how to play around with the
> 'lambda', 'sigma', and 'precision' fields of the opt structure so as to come
> up with a sufficiently accurate output in an acceptable amount of time? I
> was unable to find any detailed summary or list of practical
> guidelines/hints concerning these parameters. Furthermore, perhaps someone
> is aware of a more effective SRLS reg routine...
>
>
> Here the code I used to reshape the tri-dimensional chans x samps x trials
> matrix, run regica, and come back to the cleaned 3 dims matrix. As you see,
> opt parameters are default, with the exception of .20 instead of .25 for
> correlation threshold, note that srls is default here.
>
>
>   EEG2DIM.data = reshape(EEG.data,
> size(EEG.data,1),size(EEG.data,3)*size(EEG.data,2));
>
>   opt.EOG = [EEG2DIM.data(1,:);EEG2DIM.data((67:69),:)]; %Fp1(1) plus
> HEOG1, HEOG2, & VEOG
>
>   opt.M=3;
>
>   opt.lambda=0.9999;
>
>   opt.sigma=0.01;
>
>   opt.prec=50;
>
>   opt.crittype = 'correlation';
>
>   opt.corthr = 20;
>
>   [EEG1] = regica(EEG2DIM.data((1:64),:),opt);
>
>   [EEG2] = reshape(EEG1, size(EEG1,1), size(EEG.data,2)/size(EEG.data,3),
> size(EEG.data,3));
>   EEG.data((1:64),:) = EEG2((1:64),:);
>
>
> Any insight would be really appreciated,
>
>
> Mahesh
>
>
>
>
>
> Mahesh M. Casiraghi
> PhD candidate - Cognitive Sciences
> Roberto Dell'Acqua Lab, University of Padova
> Pierre Jolicoeur Lab, Univesité de Montréal
> <mahesh.casiraghi at umontreal.ca>mahesh.casiraghi at umontreal.ca
>
> I have the conviction that when Physiology will be far enough advanced, the
> poet, the philosopher, and the physiologist will all understand each other.
> Claude Bernard
>
>
> _______________________________________________
> Eeglablist page: <http://sccn.ucsd.edu/eeglab/eeglabmail.html>
> http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to
> <eeglablist-unsubscribe at sccn.ucsd.edu>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>eeglablist-request at sccn.ucsd.edu
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20110406/b772b601/attachment.html>