[Eeglablist] binica

[arno]

Dear Sven,

there is indeed a small difference in the two codes, which is linked (to 
my knowledge) to the bin size. Infomax does not run on the whole data 
time range but on sub-windows and the size of these windows depends on 
the data length. In a loop, a sub-window is processed then weight matrix 
is updated, etc... The optimal theoretical bin size is 1 according to 
Tony Bell but it would be too long to compute Infomax with binsize of 1. 
The computation of binsize (for historical reasons probably due to the 
fact that different programmer dealt with these 2 codes) is not exactly 
the same in runica() and binica(). I have also test the behavior of 
Infomax for different bin sizes and there are strange artifacts 
appearing for large bin sizes (the default range is fine though). I do 
not really know about the computation of the learning rate and it might 
be slightly different too. This might explain the difference you 
observe. In our experience, these functions return close to equivalent 
results.

Another explanation for the difference might be the fact that the 
shuffling of data points at each step is random so you will never obtain 
exactly the same solution.

Best,

Arno


> I have a question about binica and runica.
>
> I have installed matlab/eeglab under linux (Debian) and run binica 
> (included in eeglab 5.01b: ICA Version 1.4 Feb),  that works fine 
> (binica is quite fast :-)), but the defined weight change of  1E-7 is 
> reached much later than with runica and sometimes it isn't reached 
> even after 1000 steps  (63 chans, QuickAmp).
>
> In opposite with runica it does. Further the initial weight change is 
> very high. I used the same options  like the runica defaults, except 
> the maximum number of steps (1000). I have used the same data for both 
> and have run them several times. The problem (if it is one) seems to 
> be solved by projecting the data into a subspace by pca.
>
> So i wonder if there is a difference between runica and binica, or if 
> i made something wrong. 
> I would be deeply grateful for some suggestions.