[Eeglablist] ICA: freezing when choosing learning rate

Scott Makeig smakeig at ucsd.edu
Wed Nov 2 09:57:20 PST 2005


Elloitt -

Likely your dataset is no longer full rank (or nearly so). The weights 
keep blowing up, causing runica/binica to start over with a lower rate. 
Use rank() and cond() to check, then  try 
runica(EEG.data,'pca',EEG.nbchan-1).

Scott Makeig


Elliot Cohen wrote:

>Hi,
>
>In one of the datasets that I am trying to compute ICA for, I did several
>preprocessing steps to clean the data slightly before ICA. Namely, I removed
>any linear trend in the data, subtracted the ensemble mean (the mean of each
>time point across trials), and divided by the ensemble standard deviation.
>After this is done, I try to compute an ICA decomposition using the runica
>algorithm, the algorithm seems to get stuck trying to find a learning rate.
>This is not to say that the machine freezes, but more like it gets stuck in
>an internal loop that it isn't able to break out of. Has anyone ever
>experienced this problem? The dataset is approximately 70 trials, each trial
>has 7,144 samples, and 64 channels. I am using a 3.2 gHz machine with 1 GB
>of memory running Mandrake. I have been using eeglab for the last year or so
>to decompose data and have never run into anything like this, does anyone
>have any ideas?
>
>Thanks,
>
>Elliot
>  
>


More information about the eeglablist mailing list