[Eeglablist] Clustering / ICA sign ambiguity
s.debener at uke.uni-hamburg.de
Sun Feb 18 11:09:23 PST 2007
One way to bypass this problem is by back-projection. You are usually
interested only in a relatively small number of components/component
clusters, and those of interest could, for instance, be identified by
the amount of variance they contribute to an ERP condition effect (Plot
>Sum/compare comp. ERPs).
The new cluster functions also allow you to cluster by all sorts of
information. Some are not affected by the sign ambiguity (dipole
localization, comp power, comp time-frequency, comp ITC) and sometimes
provide valuable clustering information.
Once you have identified several components belonging to the same
cluster you could set up a little script flipping all those components
showing a polarity opposite to the one defined as a template (I usually
take the first comp from the first subject as template).
Hope this helps,
Ralf Schmälzle wrote:
>currently I am testing the clustering-program and ran into a problem
>regarding the ICA-sign-ambiguity. I have learned (and found many examples)
>that the ICA algorithm has a sign ambiguity (because of the random starting
>parameters), that is the signs of the IC-timecourse and
>scalp-projection-map are arbitrary in that only the product (=EEG.data) has
>fixed sign. This can be reached either by (ScalpMap*ICtimecourse) or by
>((-1)*ScalpMap)*((-1)*ICtimecourse) since the (-1) cancels out.
>It seems to me that this may cause a problem when it comes to clustering
>since there - according to my understanding of the clustering process - a
>sign ambiguity e.g. in the ComponentERP or CompScalpMap would cause
>corresponding Components to lie on opposite ends of the to-be-clustered
>dimension. At least when doing the grandaverage of a componentERP this
>information is essential, isn't it?
>Is that correct and how could I deal with this?
>In principle it should be allowed to flip the results (Map and TimeCourse)
>but how do I know in advance (before clustering) whether one special
>component must be flipped or not.
>Many thanks, Ralf.
>Tel. +49 (0) 7531 367501
>Ralf_Schmaelzle at web.de
> o o o o o <o <o> o> o
>.|. \|. \|/ // X \ | <| <|>
> /\ >\ /< >\ /< >\ /< >\ /<
>Fach D 36
>Raum: C 528
>Tel: +49 (0) 7531 88 2449
>Fax: +49 (0) 7531 88 2971
>Ralf.Schmaelzle at uni-konstanz.de
>eeglablist mailing list eeglablist 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
More information about the eeglablist