[Eeglablist] ICA with 128 channels

[Germán Gómez-Herrero]

Hi Maria,
As Scott pointed out, I also think that what you are observing is due to the
fact that you are not using enough data samples for ICA to "learn" the
independence of the sources underlying your data. I would add to Scott's
comment that, if you are using narrowband filters before ICA, you might need
to use even more data samples than the number recommended by Scott (see the
references below for an explanation of this). 

If you used too few data samples, ICA has probably "overlearned" your data,
and the estimated independent components represent an optimum of your ICA
contrast function only for the few data samples presented to the ICA
algorithm and do not generalize well for other than those data samples. I
would recommend you reading the following reference from J. Särela and R.
Vigario (2003), if you want to understand what is overlearning in ICA
algorithms:

http://citeseer.ist.psu.edu/682476.html 

If you are analyzing evoked potentials you might also be interested in
reading this paper by some of my colleagues and myself: 

http://www.cs.tut.fi/~gomezher/projects/eeg/cimed05.pdf 

As Scott tells you, the best solution to avoid overlearning is to use more
data samples but, if this is not an option, you can try some of the
solutions proposed in (Särela and Vigario, 2003). To rule out the
possibility that ICA has just overlearned your data can be quite tricky,
especially when performing exploratory analyses of high-dimensional
datasets. One possibility is using ICASSO:

http://www.cis.hut.fi/projects/ica/icasso/

Although the ICASSO code above includes only the FastICA algorithm, you can
trivially modify the code to use other spatial ICA algorithms (like Infomax,
JADE, etc...). If you are interested, I can send you such modified version
of ICASSO (without any guaranteed though because we are not the developers
nor the maintainers of the ICASSO toolbox). 


Regards,
Germán



 

---------------------------------------------------------------------
Germán Gómez-Herrero
M. Sc., Researcher
Tampere University of Technology
P.O. Box 553, FI-33101, Tampere, Finland
Phone:   +358 3 3115 4519
Mobile:  +358 40 5011256
Fax:     +358 3 3115 4989
http://www.cs.tut.fi/~gomezher/index.htm

From: eeglablist-bounces at sccn.ucsd.edu
[mailto:eeglablist-bounces at sccn.ucsd.edu] On Behalf Of Scott Makeig
Sent: 28 July 2008 05:14
To: mjalbrzikowski8 at ucla.edu
Cc: eeglablist at sccn.ucsd.edu
Subject: Re: [Eeglablist] ICA with 128 channels

Maria - The first question is how much data you are decomposing... e.g. With
4x the number of channels, you should use > 16x the data length in the
decomposition... For 128 channels, you should use >>128^2*20 (320k) time
points. At 250 Hz sampling rate, that would be at least 15 min or more of
continuous data or data epochs... One might say (poetically) that the
sources need this much data to express their independence...

Scott Makeig
On Sun, Jul 27, 2008 at 6:35 PM, <mjalbrzikowski8 at ucla.edu> wrote:
Hi,
  I am currently running ICA through data that was collected with a 128
channel cap.  However, unlike the data that I have previously worked with
less channels, ICA does not return identifiable blink (VEOG or HEOG)
components with this electrode setup. I'm thinking this is because after you
ICA, you get 128 components, and VEOG or HEOG is dispersed across several
different components.  However, I cannot accurately identify these
components now.  It was muich easier to pull out blink components with less
channels.  Has anyone else had this problem?  If so, did you figure out any
solutions that you would like to share?
Thank-you,
Maria Jalbrzikowski



-- 
Scott Makeig, Research Scientist and Director, Swartz Center for
Computational Neuroscience, Institute for Neural Computation, University of
California San Diego, La Jolla CA 92093-0961, http://sccn.ucsd.edu/~scott