[Eeglablist] rank inconsistency problem during ICA when using averaged mastoids reference

Fri Mar 22 15:34:03 PDT 2013

Thanks Makoto,

Yep, I've followed the discussion on the listserv; I think it's an interesting one.  I guess I was wondering what kind of workarounds people used to deal with the issue - to make the dimensions match.  (Sorry, I don't think I made that clear in my question.)  Anyway, sounds like dropping a channel is the best way to go.

Thanks,
Pete
________________________________
From: Makoto Miyakoshi [mmiyakoshi at ucsd.edu]
Sent: Friday, March 22, 2013 1:22 PM
To: Bachman, Peter
Cc: eeglablist at sccn.ucsd.edu
Subject: Re: [Eeglablist] rank inconsistency problem during ICA when using averaged mastoids reference

Dear Peter,

It's a well-known topic in this list.
Actually average referencing reduced the data rank by 1. Do you wonder why? It's because you introduce a correlation to the data by subtracting the channel average. If you still wonder or want to see mathematical explanation, you may wan to search Jason Palmer's post in the list.

So, you should discard a channel (any channel is fine) to make the number of channels matched to the actual rank.

Makoto

2013/3/21 Bachman, Peter <bachman at psych.ucla.edu<mailto:bachman at psych.ucla.edu>>
Hi everyone,

I have a question regarding a rank inconsistency problem that arises when we run ICA on a dataset that has been re-referenced to averaged mastoids.  The problem appears related to the fact that we have re-referenced offline to the average of two channels, and ICA "expects" to find 65 channels (the total number we begin with) but only finds 64.  As a consequence it defaults to using PCA to reduce the number of dimensions - something we'd prefer to avoid with these particular data.

The problem arises regardless of the ICA algorithm we use and also apparently regardless of the re-referencing parameters we use (as long as two reference channels are involved - one channel works fine).  I should also note that the data were recorded in ANT as .cnt files, but I'm guessing that's not critical to this problem.  We're using EEGLAB 11_0_5_4b.

Has anyone else run into this and found a workaround?  Or is there something else we should be doing to ensure that ICA runs, rather than PCA?  (I'm also open to the suggestion that I'm completely misdiagnosing the problem!)

I've pasted a portion of our ICA/PCA output below.  (We're using binica in this example, but the same issue arises with the regular extended infomax algorithm.)

Thanks!
Pete
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Warning: If the binary ICA function does not work, check that you have added the
binary file location (in the EEGLAB directory) to your Unix /bin directory (.cshrc file)
Warning: fixing rank computation inconsistency (63 vs 64) most likely because running under Linux 64-bit MatlabData rank (64) is smaller than the number of channels (65).
binica: using source file '.../matlab/eeglab11_0_5_4b/functions/sigprocfunc/binica.sc<http://binica.sc>'
binica(): using binary ica file '?.../matlab/eeglab11_0_5_4b/functions/resources/ica_linux'
binica(): processing 4 (flag, arg) pairs.
   setting lrate, 0.001
   setting pca, 64
   setting extended, 1
scriptfile = binica8147.sc<http://binica8147.sc>

Running ica from script file binica8147.sc<http://binica8147.sc>
   Finding 64 components.
alias erplab '.../data/erplab': Command not found.

ICA Version 1.4  (Feb. 14, 2002)

Input data size [65,754696] = 65 channels, 754696 frames.
After PCA dimension reduction,
  finding 64 ICA components using extended ICA.
PDF will be calculated initially every 1 blocks using 6000 data points.
Initial learning rate will be 0.001, block size 501.
Learning rate will be multiplied by 0.98 whenever angledelta >= 60 deg.
Training will end when wchange < 1e-07 or after 512 steps.
Online bias adjustment will be used.
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

Peter Bachman, PhD
Staglin IMHRO Center for Cognitive Neuroscience,
Center for the Assessment and Prevention of Prodromal States (CAPPS)
& Adolescent Brain and Behavior Research Clinic (ABBRC)
Semel Institute for Neuroscience and Human Behavior, UCLA
Office: (310) 206-4245<tel:%28310%29%20206-4245>
bachman at psych.ucla.edu<mailto:bachman at psych.ucla.edu>

________________________________
IMPORTANT WARNING: This email (and any attachments) is only intended for the use of the person or entity to which it is addressed, and may contain information that is privileged and confidential. You, the recipient, are obligated to maintain it in a safe, secure and confidential manner. Unauthorized redisclosure or failure to maintain confidentiality may subject you to federal and state penalties. If you are not the intended recipient, please immediately notify us by return email, and delete this message from your computer.

_______________________________________________
Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu<mailto: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<mailto:eeglablist-request at sccn.ucsd.edu>

--
Makoto Miyakoshi
JSPS Postdoctral Fellow for Research Abroad
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20130322/09762a0e/attachment.html>