[Eeglablist] Your Brain Scan Looks Different on Mac and PC

[Tim Mullen]

Since most EEGLAB functionality depends only on platform-independent native
Matlab code (i.e. no reliance on precompiled binaries/mex files), any
significant differences in the results observed between different operating
systems would have be due to core differences in Matlab itself across
different OS' and/or Matlab versions. As such, this concern would also
necessarily apply to *any* program executed in native Matlab. This would
include a fair percentage of simulation and modeling work in physics,
biology, engineering, etc. While a growing number of scientists and
engineers may be rejecting Matlab due to its high cost relative to
competitive free alternatives like Python/SciPy, I haven't heard of people
rejecting Matlab due to shoddy cross-platform replicability (although of
course some differences in results could occur between OS' and/or Matlab
versions due to differences in e.g. BLAS or LAPACK implementations, among
other things). So I don't think it's too much of a concern with EEGLAB :)

The 4e-11% difference between Mac/windows and Linux could be due a number
of things including slightly different implementations/builds of BLAS or
LAPACK libraries (which ship with Matlab and may not be strictly identical
for different OS' or Matlab release versions). In any case, a *percent
difference* this small is on the order of typical numerical imprecisions,
such as rounding errors, that one might expect to encounter (and
accumulate) in everyday computing tasks on a single machine. While one
should always be careful to ensure one's algorithm is not sensitive to this
kind of small numerical imprecision (which is why things like runica() and
Matlab filtering ops use double-precision), it's a far cry from the massive
differences found in FreeSurfer results across OS'...

Tim

On Fri, Jun 22, 2012 at 8:57 AM, Arnaud Delorme <arno at ucsd.edu> wrote:

> Dear all,
>
> some of you might have heard this distressing news: FreeSurfer, the well
> known fMRI brain imaging software, gives different results depending on the
> operating system.
>
> http://www.ncbi.nlm.nih.gov/pubmed/22675527
>
> Is it the same for Matlab? We have performed basic tests and Matlab
> numerical computation seems to be relatively OS. We have tested
>
> A - Matlab 7.8.0.347 (R2009a) on Windows XP
> B - Matlab 7.11.0.584 (R2010b) on Mac OS X (Lion)
> C - Matlab 7.14.0.739 (R2012a) on Fedora core 3.3.6-3.fc16.x86_64
>
> We have tested the ERSP time-frequency decomposition function newtimef.m
> on the EEGLAB epoched tutorial dataset ('eeglab_data_epochs_ica.set'). We
> have kept all default values.
>
> A and B were strictly identical. It might be important to note that
> Windows XP was running on top of Mac OSX in a WMWARE virtual machine so the
> same Intel processor was being used in both cases. The difference between
> A/B and C was 4e-11% on average. Still it was not strictly equivalent. This
> type of difference might become more important for algorithms such as ICA
> which are iterative. We had noticed for example that ICA should not be run
> in simple precision or the ICA components would be qualitatively much
> poorer (EEGLAB forces double precision on the data before running ICA). Any
> comments or further tests welcomed.
> Thanks,
>
> Arno
> _______________________________________________
> 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
>



-- 
---------  αντίληψη -----------
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20120622/eaac2518/attachment.html>