[Eeglablist] ICA runs slowly and returns complex numbers

Scott Makeig smakeig at ucsd.edu
Mon Dec 16 11:26:32 PST 2013


Ilana -

Since your derived bipolar channels are all still linear sums of EEG
sources, ICA will separate the data into the same sources as for the
original data.  However, their projections to the bipolar channels will be
computed by ICA, not the projections to the original channels. That means
that the 'scalp maps' returned by ICA in the columns of 'EEG.icawinv' may
not be interpretable as scalp maps (since the different bipolar channels
have unknown offsets from each other), and cannot be source localized.

I would suggest decomposing the original data with common reference, then
converting to your bipolar montage for whatever purposes you choose.  Using
the EEGLAB re-referencing tool to do this should properly convert the ICA
decomposition as well. Be sure to use a large enough data matrix for the
decomposition. ICA will be learning 64x64 = 4k weights, so the number of
time points should best be ~5-30+ times this....

Scott Makeig


On Wed, Dec 11, 2013 at 12:02 AM, Ilana Podlipsky <ilana.mlist at gmail.com>wrote:

> Just to clarify... The data is recorded in referential montage with
> reference in Cz and then converted to virtual bipolar montage, where each
> electrode is subtracted from its neighbor along the longitudinal and
> transverse axis. From 64 referential electrodes this produces 85 (bipolar)
> channels.
>
>
> On Wed, Dec 11, 2013 at 8:48 AM, Arnaud Delorme <arno at ucsd.edu> wrote:
>
>> Dear Llana,
>>
>> I do not have any reference for this. It is hard for me to understand why
>> it would work well. Maybe Jason can step in and enlighten us.
>> Best,
>>
>> Arno
>>
>> On Dec 10, 2013, at 10:45 PM, Ilana Podlipsky <ilana.mlist at gmail.com>
>> wrote:
>>
>> Dear Arno,
>>
>> Thank you for your input.
>> Can you elaborate why ICA on bipolar montage is not good? Can you give me
>> a reference for this?
>> From our experience it works quite well and separates well into
>> reasonable components of signal and artifact.
>>
>> Thank you
>> Ilana
>>
>>
>> On Tue, Nov 26, 2013 at 8:23 AM, Arnaud Delorme <arno at ucsd.edu> wrote:
>>
>>> Dear Llana,
>>>
>>> if ICA detects a rank of 57, then this is probably what you should use.
>>> ICA on bipolar montage will not be informative. ICA will attempt to
>>> model each channel reference so that common sources may be projected to all
>>> channels in a linear fashion.
>>> Best,
>>>
>>> Arno
>>>
>>> On Nov 24, 2013, at 3:18 AM, Ilana Podlipsky <ilana.mlist at gmail.com>
>>> wrote:
>>>
>>> Dear Arno,
>>>
>>> Judging by the ICA output in Matlab the solution converges but very very
>>> slowly.
>>> I use bipolar montage, each electrode is referenced to its neighbor. I
>>> use 64 electrodes recorded with one reference electrode and convert  it to
>>> 85 differential channels off line. ICA detects rank 57 but I changed it
>>> manually to 64 since this is the original number of channels, is that
>>> correct?
>>> Meanwhile I tried binica on the same computer running Ubuntu, and it
>>> runs much faster (in 20 minutes). I would still like to resolve this issue
>>> because I'd prefer to work on Windows.
>>>
>>> Thanks,
>>> Ilana
>>>
>>>
>>>
>>> On Wed, Nov 20, 2013 at 6:59 PM, Arnaud Delorme <arno at ucsd.edu> wrote:
>>>
>>>> Dear Ilana,
>>>>
>>>> did your ICA solution converge (meaning that the weight difference
>>>> decrease with time). This might be the issue.
>>>> Also, are you using average reference or linked mastoid. In this case,
>>>> the data matrix rank is the number of channels minus 1. ICA tries to detect
>>>> this automatically but sometimes fails. You then have to manually reduce
>>>> the number of dimension by 1 when running ICA. If you have 64 channels, in
>>>> the edit box for running ICA (where there is already 'extended', 1) you may
>>>> add 'pca', 63.
>>>>
>>>> Best,
>>>>
>>>> Arno
>>>>
>>>> On Nov 12, 2013, at 11:54 PM, Ilana Podlipsky <ilana.mlist at gmail.com>
>>>> wrote:
>>>>
>>>> > Hi All,
>>>> >
>>>> > Since I've recently changed my computer ICA in eeglab runs very very
>>>> slowly and returns complex numbers.
>>>> > On my previous computer, on the same data I ran the same ICA  within
>>>> an hour or two. On the new computer the same ICA takes more than 24 hours,
>>>> After 512 steps it returns this message :
>>>> >
>>>> > Sorting components in descending order of mean projected variance ...
>>>> > Warning: Matrix is close to singular or badly scaled.
>>>> >          Results may be inaccurate. RCOND = 6.956943e-019.
>>>> >
>>>> > When I try to plot the ICA activations I don't see any traces and
>>>> when I look into the EEG.icaact matrix I see only complex numbers. Tried
>>>> both runica and binica.
>>>> > This has never happen to me with the old computer on the same dataset.
>>>> > Both the old and new computer run Win7 64bit, matlab 2008a and eeglab
>>>> 12. The hardware of the computers is different.
>>>> >
>>>> > What could be the reason for this, and what can ?I do to solve this?
>>>> >
>>>> > Thanks for the help,
>>>> > Ilana
>>>> > _______________________________________________
>>>> > 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
>>>>
>>>>
>>>
>>>
>>
>>
>
> _______________________________________________
> 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/20131216/ba620d3d/attachment-0001.html 


More information about the eeglablist mailing list