[Eeglablist] 64 vs 128

[Scott Makeig]

Matthew -

We are collecting 256 channels and running binary ICA, etc. on Opteron 
Linux nodes with 8 GB RAM. Using 64-bit Matlab, we can compute on 
matrices of > 4 GB.  However, we find we need at least on the order of 
30*256^2 time points to learn the 256^2 weights. Decomposing 128 
channels, meanwhile, is straightforward on this system, as is computing 
time/frequency transforms, etc. With 128 channels, you can collect data 
on the sides of the head and forehead. We cannot vouch for the benefit 
of 256 over 128 at this point (and have not focused on answering this 
question), but expect that 128 can be more informative than 64. We are 
quite satisfied with our Biosemi active electrode system and estimate 
that two techs could put on a 128-channel montage in 15 mins. Our own 
designed full-head cap, at least, is quite comfortable as well. I do not 
know the latest responses of competitors to the Biosemi design, however. 
Best wishes!

Scott

Subject:
ICA decomposition with 128 versus 64 channels?
From:
mkb30 at cam.ac.uk (Matthew Belmonte)
Date:
Wed, 18 Jan 2006 23:41:05 +0000

To:
eeglablist at sccn.ucsd.edu


I'm in the process of putting together a proposal for an EEG facility, and
would like to select hardware with EEGLAB processing in mind.  I'm approaching
this from perhaps a bit of a dated perspective: in 1996 I was using only 16
channels and homebrewed software for time-frequency analysis, and I've spent
the intervening decade working exclusively with fMRI.

I've heard from one EEGLAB user that 128 channels don't confer much advantage
over 64, since inputs must be spatially downsampled in order to be processed
practically on typical computing hardware, and since the independent components
of interest (those from neural sources) don't become much cleaner with 128
inputs as compared to 64.  (The tradeoff of spatial resolution and SNR to
electrode application time also is a consideration; we'd be recording from
autistic children and couldn't afford any great deal of time spent fiddling.)

I'd like to hear from EEGLAB users (and developers!) with experience at 128
and/or 64 channels:  Do you find a 64-channel system adequate?  What
improvement in data quality has moving to 128 channels given you?  If I loaded
up a GNU/Linux system with the most RAM that I could get (16GB on an IBM
IntelliStation), would it be able to handle an ICA decomposition of 128-channel
data without thrashing, or would I be doubling my investment in amplifiers only
to have to mix down 128 signals to 64 before ICA?  And, even if it would be
computationally practical, would it be scientifically useful enough to justify
the extra preparation time?

Many thanks

Matthew Belmonte <mkb30 at cam.ac.uk>