sourcestatistics

Robert Oostenveld r.oostenveld at FCDONDERS.RU.NL
Mon May 9 08:56:45 CEST 2005


Hi Masao,

On 6 May 2005, at 20:34, Matsuhashi, Masao (NIH/NINDS) wrote:
> I have questions about statistical measures using random re-sampling
> and
> permutation.
>
> i)	Where can I find the sourcestatistics.m? According to the tutorials
> or documents, this script seems essential but I couldn't find it in
> recent
> distributions of FieldTrip.

I thought that it was included in the 0.9.6 version, but it seems not
to. I'll send it to you in a separate mail. We are about to release
1.0, which will include it.

> ii)	In the sourceanalysis.m, re-sampling or permutation uses lots of
> memory and the computation ends up with OUT OF MEMORY. (e.g. rnd_aCf
> and
> rnd_bCf use 2 conditions * 500 trials * 274 channels * 274 channels *
> 16
> bytes/complex = 1.2GBytes)
> Is there any way to go around this problem, such as repeating 100 trial
> random re-sampling for 10 times, or must I switch to Opteron?

There are two memory consuming aspects: first, each randomization takes
uses two cross-spectral density matrices (one for each condition). The
second memory consuming aspect is the source reconstruction volumes, in
your case that would consist of (500+1)*2 volumes. Each volume contains
of several parameters (depending on what you are beaming: pow, noise,
coh), increasing the number of volumes with a factor of 2 or 3. The
size in bytes of a volume depends on the resolution that you choose,
but I suggest that you also try to estimate whether that would fit into
memory prior to starting the computation. Btw. 500 is quite a lot, we
typically find that the probabilities don't change that much after
200-300.

You can indeed run multiple randomizations separately. If you do that
in separate matlab sessions (on different machines), make sure that you
use the matlab seed() function to prevent the random number generator
from creating the same random sequence in every matlab session.
Afterwards, you have to concatenate the contents of the source.trialA
and trialB fields.

> iii)	As expected, this is very time-consuming process. Can I somehow
> use
> the Beowulf support feature, and if yes, how? We have Beowulf cluster
> available, but the number of Matlab license is limited to 8....

It is very timeconsuming indeed. Are you using precomputed leadfields?
And a decent grid resolution (not too high)? I have been trying around
with different paralellization algorithms, and also found that the
number of Matlab licenses was more of a problem than the number of
machines. The paralellizationis still in an experimental stage. The
latest and most promising implementation of the parallellization (the
"mentat" version) code can be fuond as a separate toolbox on
http://oase.uci.kun.nl/~roberto/mentat/. It uses only one matlab
license on the head node and creates standalone binaries for the slave
nodes using the matlab compiler. If you want, you can give it a try and
see whether you can get it to work on your cluster.

best regards,
Robert


----------------------------------------------------------------------
Robert Oostenveld, PhD
F.C. Donders Centre for Cognitive Neuroimaging
Radboud University Nijmegen
phone: +31-24-3619695
http://www.ru.nl/fcdonders/
----------------------------------------------------------------------
N.B. Starting from 1 September 2004, the University of Nijmegen has
changed its name to Radboud University Nijmegen. All web- and
email-addresses ending in ".kun.nl" should therefore be changed into
".ru.nl". Please update your address book and links.



More information about the fieldtrip mailing list