[Eeglablist] Questions on FFT phase results and ICA methodology

Wed Jul 18 16:17:39 PDT 2018

Dear Chris,

> is there a way to have runica(or any of the other ICA functions) process
the timeline in a sliding fashion every specific number of epochs using
command inputs instead?

Usually, such a sliding-window solution is used to address non-stationarity
of data. See my colleagues' works.
https://github.com/goodshawn12/REST
https://ieeexplore.ieee.org/document/6533850/

AMICA is also capable of learning the specified numbers of models (i.e.,
ICA weight matrices) from data in an unsupervised way.

In this case, however, you'll have a lot of decompositions (another
colleague of mine used it for dictionary learning). You need to know what
you want to do with the overwhelming amount of data. Our standard EEGLAB
process rather recommends not to split ICA across conditions wherever
possible so that subsequent statistical test is within IC test (like
performing paired t-test). Good thing about it is that the consistency of
source locations etc between the conditions is guaranteed. From the
viewpoint of data learning, it is also a good idea to feed ICA as long data
as possible.

> Didn't see anything in the help files that indicated this was the case,
but it seemed like a cleaner solution.

Generally, standard researcher's goal is not to have ever evolving ICA
weight matrices across time. Usually it is not very helpful because it is
against reductionistic principle in data mining. But if you are doing
research on data nonstaionarity, you need a solution to represent
nonstationarity in one way or the other.

Makoto

On Wed, Jul 18, 2018 at 2:44 PM Chris Rose <u6t9n7 at u.northwestern.edu>
wrote:

> Hi Makoto,
>
> Thanks for the response! I'll play around with the newtimef info you sent,
> and may have a couple followup questions.
>
> As for the ICA, skipping generation of the additional ICA statistics made
> a huge difference in bringing down the file size. I've also been handling
> the sliding windows somewhat crudely, just breaking the necessary segment
> for each window into a new set and processing the ICA from there; however,
> the data is already processed through regepoch with 1s intervals, is there
> a way to have runica(or any of the other ICA functions) process the
> timeline in a sliding fashion every specific number of epochs using command
> inputs instead? Didn't see anything in the help files that indicated this
> was the case, but it seemed like a cleaner solution.
>
> Best,
>
> Chris
>
> On 17 July 2018 at 16:46, Makoto Miyakoshi <mmiyakoshi at ucsd.edu> wrote:
>
>> Dear Chris,
>>
>> > however I'm unclear on how the post-newtimef calculation would look.
>>
>> Although not described in newtimef() help, if you see the function line
>> 369,
>>
>> [P,R,mbase,timesout,freqs,Pboot,Rboot,alltfX,PA] = newtimef( data,
>> frames, tlimits, Fs, varwin, varargin);
>>
>> The 8th output alltfX is a 3-D tensor of complex values, namely freq x
>> time x trials. To calculate single-trial phase, simply angle(alltfX).
>>
>> > 2) A little more unorthodox, I was hoping to include ICA information
>> for 1-minute epochs across the entire recording to see if there are changes
>> in the predominant components as the stimulus progresses. Trying to simply
>> segment an EEG file into minutes and run ICA on each segment resulted in a
>> truly enormous amount of data when testing it with just one subject
>> (~60GB), and that would certainly be prohibitive to including it in our
>> analysis. Is there any way that this might be done more efficiently such
>> that the resulting data is of a mroe manageable size?
>>
>> Even if you have 256 ch data for 240 sliding windows, the total size of
>> the weight matrix is double(256 x 256 x 240) = 126MB per subject. You need
>> to generate and save this matrix for EEG.icaweights and EEG.icasphere, so
>> you need 126 x 2 = 252 MB. I don't know why you say you need 60GB, unless
>> you have 240 subjects x 252 MB = 60GB in which case you have no choice.
>>
>> If you haven't done so, consider saving only EEG.icaweights and
>> EEG.icasphere for each iteration.
>>
>> Makoto
>>
>> On Sun, Jul 15, 2018 at 9:11 PM Chris Rose <u6t9n7 at u.northwestern.edu>
>> wrote:
>>
>>> Hi all,
>>>
>>> I am working on a project where we'll be using A.I. to do an exploratory
>>> analysis on recordings taking during the viewing of a stimulus that lasted
>>> ~90minutes. Given the length of the recording period and the number of
>>> subjects(60), we are trying to be selective about what information we use
>>> as learning inputs to avoid overloading the AI algorithm.
>>>
>>> There are two things that I was hoping to include as inputs in the AI
>>> analysis that I'm having trouble getting.
>>>
>>> 1) I am hoping to include phase calculations taken from FFT of the
>>> standard frequency bands. We've been using spectopo to calculate the power
>>> in each band, but I don't see any option to calculate phase with this
>>> function. This thread
>>> <https://sccn.ucsd.edu/pipermail/eeglablist/2011/003837.html> from a
>>> previous eeglablist question seems to have a similar question, and the
>>> original writer says they solved their problem by using newtimef outputs to
>>> calculate phase for each band/channel, however I'm unclear on how the
>>> post-newtimef calculation would look. Since the data set is so large, I'm
>>> hoping to avoid actually calculating coherence prior to the AI, and simply
>>> let the neural net calculate it internally.
>>>
>>> 2) A little more unorthodox, I was hoping to include ICA information for
>>> 1-minute epochs across the entire recording to see if there are changes in
>>> the predominant components as the stimulus progresses. Trying to simply
>>> segment an EEG file into minutes and run ICA on each segment resulted in a
>>> truly enormous amount of data when testing it with just one subject
>>> (~60GB), and that would certainly be prohibitive to including it in our
>>> analysis. Is there any way that this might be done more efficiently such
>>> that the resulting data is of a mroe manageable size?
>>>
>>> Thanks in advance for any thoughts you can offer with either issue!
>>>
>>> Best,
>>>
>>> Chris
>>> _______________________________________________
>>> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
>>> To unsubscribe, send an empty email to
>>> eeglablist-unsubscribe at sccn.ucsd.edu
>>> For digest mode, send an email with the subject "set digest mime" to
>>> eeglablist-request at sccn.ucsd.edu
>>
>>
>>
>> --
>> Makoto Miyakoshi
>> Swartz Center for Computational Neuroscience
>> Institute for Neural Computation, University of California San Diego
>>
>
>

-- 
Makoto Miyakoshi
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20180718/01c3df08/attachment.html>