[Eeglablist] ICA weights blowing up

[Scott Makeig]

Alana -

This problem is quite likely a result of trying to decompose 'rank
deficient' data, caused by two ore more channels being identical or linearly
related to one another. A common example is retaining an all-0 channels, or
following conversion to linked ears (A+B) reference, retaining e.g. A-B,
B-A, and B in the dataset as separate channels. However, these 3 channels
are not independent, since any two can be linearly combined to produce the
third. To solve this problem, either remove one (or more) of the redundant
channels from the dataset before decomposition, or reduce the dimension of
the dataset by one (or as necessary) using PCA.

Scott

On Tue, Jul 29, 2008 at 12:00 PM, Alana Firl <ajfirl at gmail.com> wrote:

> Hi, my lab has been using ICA to identify blinks and other artifacts.
> However I've found that about 20% of subjects appear to have data that ICA
> can't find a good solution for. runica takes hours to do just a few steps
> and eventually matlab crashes. All of our data after acquisition is
> processed in eeglab. The data is 128 channels, which we reduce to 75 using
> PCA. After premilinary artifact rejection, each subject has about 30 minutes
> worth of epoched data. I use extended runica, although other versions of ica
> didn't do any better. I did some troubleshooting with the data and runica.m
> and it seems as though, for the "bad" subjects, the weights are blowing up
> every time. For some subjects, their data is just poor quality but for
> others, their data looks clean, and I don't know why ICA isn't working for
> them. Is there anything I can do to get around this problem?
> Thanks!
> Alana Firl
> UC Davis
> Imaging Research Center
>
> _______________________________________________
> 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
>



-- 
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
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20080730/c6121fe4/attachment.html>