# [Eeglablist] Average Reference

Makoto Miyakoshi mmiyakoshi at ucsd.edu
Thu Mar 26 10:45:09 PDT 2015

```Dear Claudio,

First of all I'm not qualified to discuss it since my background is not so
mathematical. That being said, let me tell you what I know.

> But if there are not 64 (on a 64 electrode setup) unique brain-related
sources in the data, the ICA tends to split one source into two,

That's not the case. ICA by nature performs independent subspace
decomposition. Our current interpretation of such IC subspace is that if
there is a moving source ICA may decompose it into a subspace (imagine you
take still picture of a moving car...)

> (they are independent by means of the ICA, but occur on the same instant
in the brain, thus having a entropy which deviates from zero).

If you measure mutual information reduction (MIR) across the ICs, you'll
see the subspace. I wonder, by the way, 'getMIR.m' is available outside
SCCN... it's packaged in AMICA plugin.

> And what would be the solution to the problem.

I heard people explain it as if it is not a problem since it's a nature of
ICA... but I'm still puzzled, yes!

> I normally try to give the most cleaniest data into the ICA computation
(excluding noisy channels, eye movement correction, raw data inspection
etc.), so shouldn't i do that?

Our clustering approach, either k-means or MeasureProjection, addresses the
issue. However, if you need to choose one out of several ICs, you may want
to use some criteria such as largest variance, or most similar
spectrum/scalp map/ERP to the cluster centroid, etc... there is no
established way to do it unfortunately. It is a weak point of ICA; it
messes up later processes.

Makoto

On Mon, Mar 23, 2015 at 2:38 AM, Claudio Georgii <
Claudio.Georgii at stud.sbg.ac.at> wrote:

> Dear Makoto,
>
> Thank you for your help and advices. I never thought that the position of
> cap would be source of the problem, but i think you are right, and it's a
> simple solution to the problem as well!
>
> I watched the videos of the link you sent me. But i am not quite sure if i
> get your line of argumentation right.
>
> Normally each artificial source is represented in one ICA component, thus
> having cleaner results in having "only" brain-related ICA components. But
> if there are not 64 (on a 64 electrode setup) unique brain-related sources
> in the data, the ICA tends to split one source into two, which form an ICA
> subspace and are residual dependent of each other (they are independent by
> means of the ICA, but occur on the same instant in the brain, thus having a
> entropy which deviates from zero).
>
> Did i got this right? And what would be the solution to the problem. I
> normally try to give the most cleaniest data into the ICA computation
> (excluding noisy channels, eye movement correction, raw data inspection
> etc.), so shouldn't i do that? Or is there a midway between clean and dirty
> data i should take.
>
> What would be your recommendations?
>
> Thanks in advance!
>
> Best,
> Claudio
>
> 2015-03-21 0:56 GMT+01:00 Makoto Miyakoshi <mmiyakoshi at ucsd.edu>:
>
>> Dear Claudio,
>>
>> > Thus, based on the theoritical assumptions of the average reference the
>> summed activity of all electrodes would deviate from zero (more than it
>> normally does), resulting in a situation, where partial activity of frontal
>> sources is substracted from all electrodes.
>>
>> I think that is true.
>>
>> > In theory this might lead sources migrate to more posterior sides?
>>
>> No, that does not happen. If you see the effect of average referencing
>> for each frame, it just adds constant values for all the channels.
>> Therefore, it does not change the inter-channel relations. If a source
>> moves, it should change the inter-channel relations.
>>
>> > In fact, since quite some time we are wondering why the effects in our
>> experiments are emerging on posterior sides than reported in the literature.
>>
>> If you apply EEG cap wrongly it may happen. Make sure that your Fp1 and
>> Fp2 are right above the eyebrows.
>>
>> > In theory i'd say that this is "bad luck" and the ICA is splitting one
>> dipole into two different ICs.
>>
>> The more channels and the clearner data you have, the more subspace
>> you'll see after ICA. For the nature of subspace, see the following video
>> by Jason Palmer.
>>
>> Chapter 2 Part 2 (Don't miss the spectacular movie from 13:01)
>> http://sccn.ucsd.edu/eeglab/Online_EEGLAB_Workshop/EEGLAB12_ica1.html
>>
>> Makoto
>>
>> On Fri, Mar 20, 2015 at 2:47 AM, Claudio Georgii <
>> Claudio.Georgii at stud.sbg.ac.at> wrote:
>>
>>> Dear Makoto,
>>>
>>> thanks for your reply! It's good to know that using 64 electrodes should
>>> be no concern if you are using an online average reference system.
>>>
>>> I now analyzed the data using the MPT (within EEGlab) and i know have 2
>>> other concerns:
>>>
>>> 1) The electrode montage we use is not evenly distributed over the head
>>> having more electrodes at posterior sites (PO9, PO10, Oz) than on frontal
>>> sides. I now asked myself the question, whether this may lead to
>>> topographic changes, and thus migration of sources (compared to 64
>>> Electrodes evenly distributed over the head), if you are using an online
>>> average reference system. I thought of this, because the fact that i have
>>> lesser electrodes on frontal sides would mean that i am not that capable of
>>> measuring the activity from all the underlying frontal sources. Thus, based
>>> on the theoritical assumptions of the average reference the summed activity
>>> of all electrodes would deviate from zero (more than it normally does),
>>> resulting in a situation, where partial activity of frontal sources is
>>> substracted from all electrodes. In theory this might lead sources migrate
>>> to more posterior sides?
>>>
>>> In fact, since quite some time we are wondering why the effects in our
>>> experiments are emerging on posterior sides than reported in the literature.
>>>
>>> 2) The ICA (infomax, extended, runica function) i used creates to almost
>>> identical ICs, which just differentiate in their respective polarity,
>>> located at the center of the brain. In theory i'd say that this is "bad
>>> luck" and the ICA is splitting one dipole into two different ICs. But the
>>> problem is that almost every participant in our study has these two ICs and
>>> based on the MPT they are later clustered into one source! Which is of
>>> course not wrong, but it inflates the dimension of the source i got from
>>> the MPT, because two dipoles (related two the two ICs) are used to cluster
>>> the source instead of one.
>>>
>>> Is there any way to prevent the ICA in splitting this one dipole into
>>> two IC's?
>>>
>>> Thanks for your help in advance!
>>>
>>> Best,
>>> Claudio
>>>
>>> 2015-03-19 20:10 GMT+01:00 Makoto Miyakoshi <mmiyakoshi at ucsd.edu>:
>>>
>>>> Dear Claudio,
>>>>
>>>> > As far as i know by now, different reference electrodes do not affect
>>>> the scalp distribution.
>>>>
>>>> Oh it does change the scalp distribution. For ICA it does not.
>>>>
>>>> > I asked me this question, because theoretically the summed activity
>>>> of all electrodes equals zero only if you have a comprehensive electrode
>>>> montage.
>>>>
>>>> That's a good point. And the head should be a perfect sphere (with no
>>>> neck) too :-)
>>>>
>>>> > Thus, are 64 electrodes enough for using an online average reference?
>>>>
>>>> Practically no problem.
>>>>
>>>> Makoto
>>>>
>>>> On Fri, Mar 13, 2015 at 6:12 AM, Claudio Georgii <
>>>> Claudio.Georgii at stud.sbg.ac.at> wrote:
>>>>
>>>>> Hi,
>>>>> i am Claudio Georgii, a Ph.D. student at the University of Salzburg,
>>>>> Austria. I am currently working with EEGlab to analyse my EEG data in the
>>>>> source space using the measure projection toolbox (MPT).
>>>>>
>>>>> I try do evaluate know whether using an online average reference (64
>>>>> electrode system) has an impact on the source distribution. As far as i
>>>>> know by now, different reference electrodes do not affect the scalp
>>>>> distribution. But have different electrode montages, while using an online
>>>>> average reference, an impact on the scalp distribution?
>>>>>
>>>>> I asked me this question, because theoretically the summed activity of
>>>>> all electrodes equals zero only if you have a comprehensive electrode
>>>>> montage. Thus, are 64 electrodes enough for using an online average
>>>>> reference?
>>>>>
>>>>> Best,
>>>>> Claudio
>>>>>
>>>>> _______________________________________________
>>>>> 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
>>>>>
>>>>
>>>>
>>>>
>>>> --
>>>> Makoto Miyakoshi
>>>> Swartz Center for Computational Neuroscience
>>>> Institute for Neural Computation, University of California San Diego
>>>>
>>>
>>>
>>
>>
>> --
>> Makoto Miyakoshi
>> Swartz Center for Computational Neuroscience
>> Institute for Neural Computation, University of California San Diego
>>
>
>

--
Makoto Miyakoshi
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/20150326/a8eb62b6/attachment.html>
```

More information about the eeglablist mailing list