AMICA fails to converge

[jvdc8]

Hi,

I am running AMICA1.7 through eeglab2024.2 on Matlab 2023B to decompose 1-second epochs of EEG data (size(EEG.data): 116 x 256 x 1000). In a first attempt with the default number of iterations set at 2000 ('pca' =115), AMICA failed to converge and stopped after reaching 2000 trials (as instructed). Increasing the number of max. iterations to 10000 results seems affects the duration of a single iteration (it increases) in the AMICA algorithm as well. I was wondering if this is expected behavior? Does the lack of convergence also imply that something may be wrong with the data?

Kind regards,
Joris

[japalmer29]

Hi Joris, The iteration time shouldn’t increase with the max iterations. I’m not sure why that would happen. Only the LL output increases which is written every ‘writestep’ iterations, but even at writestep of 1 it shouldn’t really increase iteration time going from 2000 to 10000 length. It may be that there is noise in the data causing the lrate (learning rate) to be decreased so small that it doesn’t really move. Check the out.txt to see the lrate over iterations. I would check the data for large artifact periods. You can do a quick check doing >>figure, plot(EEG.data(:,:)’), and look for large spikes in amplitude. You can’t always see such periods looking at the data in eeglab scroll plot. Or try rejecting epochs by inspection. If the lrate stays constant at 0.5 or maybe 0.25 or 0.125, it should converge such that the nd (norm of the delta weights) is less than 0.00005 and LL is increasing in the 8th or 9th decimal place. It usually doesn’t reach the default stopping conditions for nd or LL change, but if the nd is small and the LL changing in 8th or 9th decimal there probably won’t be any more detectable changes. I usually run it for 2500 max iterations by default. Best, Jason From: jvdc8 ***@***.***> Sent: Thursday, November 28, 2024 7:24 AM To: sccn/amica ***@***.***> Cc: Subscribed ***@***.***> Subject: [sccn/amica] AMICA fails to converge (Issue #48) Hi, I am running AMICA1.7 through eeglab2024.2 on Matlab 2023B to decompose 1-second epochs of EEG data (size(EEG.data): 116 x 256 x 1000). In a first attempt with the default number of iterations set at 2000 ('pca' =115), AMICA failed to converge and stopped after reaching 2000 trials (as instructed). Increasing the number of max. iterations to 10000 results seems affects the duration of a single iteration (it increases) in the AMICA algorithm as well. I was wondering if this is expected behavior? Does the lack of convergence also imply that something may be wrong with the data? Kind regards, Joris — Reply to this email directly, view it on GitHub <#48> , or unsubscribe <https://github.com/notifications/unsubscribe-auth/ACRBESQAVXK5Y6OMAOI3CTD2C4DO7AVCNFSM6AAAAABSVARQSKVHI2DSMVQWIX3LMV43ASLTON2WKOZSG4YDCOJZGU2TGNQ> . You are receiving this because you are subscribed to this thread. <https://github.com/notifications/beacon/ACRBESSFUOTAEUNFTA7COYD2C4DO7A5CNFSM6AAAAABSVARQSKWGG33NNVSW45C7OR4XAZNFJFZXG5LFVJRW63LNMVXHIX3JMTHKCDJOCA.gif> Message ID: ***@***.*** ***@***.***> >