Fwd: [Eeglablist] noise robustness of ICA

[Ronald Phlypo]

In addition to the previous submission to the list I would draw your attention
on the units in ordinates. The values that were plotted were the Frobenius norm
from the difference matrix (clean_simulated_EEG - cleaned_EEG), where
cleaned_EEG is the result from removing the artifact making use of a method
(e.g. runICA). The dip thus represents thus an increase in accuracy for certain
noise levels? Would ICA prefer some amount of noise in the data to ameliorate
its results?



----- Doorgestuurd bericht van Phlypo Ronald <ronald.phlypo at UGent.be> -----
    Datum: Tue, 17 Jan 2006 18:34:19 +0100
      Van: Phlypo Ronald <ronald.phlypo at UGent.be>
Antwoorden aan:Phlypo Ronald <ronald.phlypo at UGent.be>
 Onderwerp: [Eeglablist] noise robustness of ICA
      Aan: eeglablist at sccn.ucsd.edu

Dear Mister/Miss,

I’m currently working on removing eye artifacts in real time applications. I
wanted to make a study of the noise robustness and I encountered a strange
dip in the graphs. We discussed the set-up among colleagues, and it seems
ok, except the results. Is there anyone of you who has some idea how to
interpret the results? Does ICA has a preference for certain noise levels?

Thanks in advance,

Ronald

PS: please find figures and simulation set-up description in attachment

Phlypo Ronald

Ronald.Phlypo at UGent.be
UGent - Faculty of Engineering
Electronics and Information Systems
Medical Image & Signal Processing
St. Pietersnieuwstraat 41
B-9000 Ghent
Belgium

tel.         +32 (9) 240 43 85
mobile   +32 (485) 05 97 81
fax.        +32 (9) 240 49 71
-------------- next part --------------
A non-text attachment was scrubbed...
Name: setup.doc
Type: application/msword
Size: 27136 bytes
Desc: not available
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20060118/cc1a0c9c/attachment.doc>
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: unnamed
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20060118/cc1a0c9c/attachment.ksh>