[Eeglablist] Regarding slow ICA

[Julian Yaoan Cheng]

Hi Narendra,


That error message doesn't tell you much, as it is always reported the first time the 'state' flag is passed to the random family of functions and just serves as a reminder to warn people before they axe the functionality for real.


The more telling message is the "Lowering learning rate to X and starting again". This can indicate that the function is having a hard time trying to even start, and is usually associated with problems in the data. While I'm not sure exactly what it means (poor data quality maybe?), I found that the only way to avoid it is through preprocessing. Downsample the data to something like 250Hz if you have a much higher sampling frequency, or try different filter settings. Makoto's preprocessing suggestions can really help here (look it up on the wiki).


Best,

Julian

________________________________
From: eeglablist <eeglablist-bounces at sccn.ucsd.edu> on behalf of Narendra Kumar <narendra.linguistics at gmail.com>
Sent: Saturday, August 26, 2017 7:53:04 AM
To: eeglablist at sccn.ucsd.edu
Subject: Re: [Eeglablist] Regarding slow ICA

I observed the differences between the two in command window as follows:

When files run in 3-3.5 hrs
Beginning ICA training ... first training step may be slow ...
step 1 - lrate 0.00100 wchange ***** angledelta 0.0 deg
step 2 - ......

When files take more than 20 hours
Beginning ICA training ... first training step may be slow ...
Warning: Using 'state' to set RAND's internal state causes RAND, RANDI, and RANDN to use legacy random number generators.  This syntax will be
removed in a future release.  See Updating Your Random Number Generator Syntax to use RNG to replace the old syntax.
> In runica at 830
  In pop_runica at 419
  In preprocessing_1_30 at 131
Lowering learning rate to 0.0009 and starting again.
Lowering  learning rate to .....

​I think the error can tell about it in a better way.

Regards,
Narendra​



On Sat, Aug 26, 2017 at 8:14 PM, Narendra Kumar <narendra.linguistics at gmail.com<mailto:narendra.linguistics at gmail.com>> wrote:
Dear EEGLAB List users,

I run ICA on 25 subjects data using script for multiple subjects data (pop_runica(EEG, 'extended', 1, 'interupt', 'off')). Till 13 subjects, it took 3.5 hours for each data file, but after then the next data file has been running for last more than 20 hours.

Kindly tell me the reason for this and suggest a solution ? I'll be thankful to you for this.

Regards,
Narendra

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20170828/0ea588d7/attachment.html>