[Eeglablist] How to correctly break down AR runica() in case of huge sets.

[Jason Palmer]

Dear Mahesh,

 

I would suggest that you run ICA on contiguous blocks of data, as in your
prototypical situation (1), or e.g. 3 ICAs on conditions 1+2, 3, and 4+5.
The ICA model expects the data to be stationary, and EEG data should be more
stationary over contiguous data. The tradeoff is between having enough data
to get good components with ICA, and having too much data so that the data
cannot be modeled as being generated by a single complete basis set. So I
suggest you divide the data into segments with e.g. (nchan*5000) samples in
each segment, and run ICA on each segment, then try to match the components
you are interested in in the different decompositions, and back-project
those components using the separately determined activations.

 

Regardless of how you divide up the data, as long as there is enough data in
each segment to determine the components present (e.g. eye-blink), ICA
should be able in principle to remove all of the variance due to the
artifact. In your second example, if 20 trials of condition 3 contain enough
data to determine the eye-blink component, then you should be able to
separately remove the eye-blink from the blink trials in each segment using
that segment’s decomposition to remove and back-project. The main issue is
having enough data to determine the artifact or non-interesting components
and remove them, and to determine interesting components, and at the same
time have the data be stationary. High-pass filtering (before epoching) is
usually necessary to remove drift and achieve basic mean stationarity.

 

So given nonstationary data, a “big global ICA” is not necessarily the best
thing, but having enough data to get good components is important. So as
long as you have enough data in your decompositions, you might even prefer
to do ICA on shorter segments, or single conditions, so that you can more
effectively remove non-stationary artifacts. Ideally you would find the type
of components you are interested in in the separate decompositions and be
able to back-project them. The power in the separately back-projected trials
should be comparable if you are back-projecting the same component(s) (e.g.
frontal midline, or mu).

 

There currently no simple way to do the swapping without modifying the code.
I will try to add such a facility in future development. We have a cluster
version of ICA at SCCN which you might be able to compile if you have a
Linux cluster. Parallelizing over multiple cores probably won’t help if your
issue is RAM.

 

Best,

Jason

 

From: Mahesh Casiraghi [mailto:mahesh.casiraghi at gmail.com] 
Sent: Monday, December 13, 2010 2:44 PM
To: japalmer29 at gmail.com; eeglablist at sccn.ucsd.edu
Subject: Re: [Eeglablist] How to correctly break down AR runica() in case of
huge sets.

 

Dear Jason,

 

thank you for all the useful hints provided along with your response.

 

I see your points, but I am still not sure on how performing ICA on
arbitrarily sampled subgroups of trials can be methodologically proof. I try
with a simple example: let's suppose we have to run an ICA for artifact
rejection in one experiment with 5 conditions with 100 trials each, and for
some reasons the physical features of the stimuli in condition 3
dramatically increase the probability of eyeblinks in those trials where
those stimuli are present. Let's consider the extreme situation in which all
the blinks would be concentrated only in 100% of trials belonging to the
experimental condition 3. Then we run, according to what reported above, 5
independent ICAs, each one on 100 trials. 

 

Now, if we do not care about the proportion of trials per condition in each
of the 5 ICAs, we can end up with two prototypical situations:

 

- one where we will have the 100 trials of the critical condition 3 all in
one ICA (and thus we will be likely to observe a big "blink component"
accounting for a lot of variance),

 

- and one where we will have 20 trials of condition 3 (and so with blinks)
in each one of the 5 ICAs, and we will therefore likely to observe 5 blink
components explaining approximately 1/5 of the big blink component above,
but this time in each one 5 ICAs.

 

The point is: as far as the sum of the accounted variances by each of the 5
components is not identical to the one accounted by the single big
component, we know that we introduce a bias in performing solution 1 rather
than solution 2. Perhaps that does not imply that this difference reflects
the amount of neural activity we erroneously removed from one condition
rather than from another, but as far as our subgroups are not balanced
condition-wise, it means that we will introduce some artifactual
condition-related variability in our data.

 

In line, the question is: should we concern about the proportion of trials
per experimental condition introduced in each of the "subgroup ICAs",
whenever we would need to decompose the ICA as a consequence of processing
constraints? 

 

And, if we do, to what extent our final backprojected data will eventually
be equal to the output of a big global ICA?

 

Third and last question: 

I see you write:

 

It would also be possible to modify the ICA algorithm to swap out data from
the disk, but as I said, I doubt using all the data would improve the
results over using as much data as you can load into memory. 

 

Is there a function (or in alternative a relatively easy way) to run runica
like that, or to run it in parallel on multiple machines - cores in a
Cluster-GPU like manner, maybe making use of the parallel processing
toolbox?

 

Hope I did not abuse of your helpfulness with all those kind of issues.

 

Cheers,

 

Mahesh

 

       

 

Mahesh M. Casiraghi

PhD candidate - Cognitive Sciences

Roberto Dell'Acqua Lab, University of Padova

Pierre Jolicoeur Lab, Univesité de Montréal

mahesh.casiraghi at umontreal.ca

 

I have the conviction that when Physiology will be far enough advanced, the
poet, the philosopher, and the physiologist will all understand each other.

Claude Bernard





On Mon, Dec 13, 2010 at 4:36 PM, Jason Palmer <japalmer29 at gmail.com> wrote:

Hi Mahesh,

 

Merging the results by simple averaging probably won’t work since the
components are returned in random order (even after the variance sorting,
components won’t necessarily have the same index.) Using matcorr() or a
similar component matching algorithm before averaging is one possibility.

 

But it seems to me that averaging will not improve anything in your
situation. As long as you have enough data in each data block that ICA runs
on, then the components you get should be well determined, allowing you to
remove the artifacts separately, and use the separate unmixing matrices to
decompose the different subsets.

 

I’m not sure what kind of analysis you’re doing, but for many purposes, you
want to identify brain components of interest and then analyze the
activations and possibly localize them. In this case you only need to match
up the components of interest in the separate decompositions, e.g. a frontal
midline ERN component, and collect all the trials with the activations
produced by the respective ICA unmixing matrices.

 

Again, as long as you use as much data as you can load (possibly overlapping
data blocks), the decompositions should be good by themselves. Comparing the
components of interest across decompositions will give you an idea of how
stable the components you’re looking at really are in your dataset. You
might also look into characterizing the variance of the component maps in a
bootstrapping sense, using a large number of resampled blocks.

 

It would also be possible to modify the ICA algorithm to swap out data from
the disk, but as I said, I doubt using all the data would improve the
results over using as much data as you can load into memory. To me it makes
more sense to verify the stability of the components you’re interested in,
and use the separate ICA unmixing/sphere matrices on their corresponding
data blocks, and separately back-project the components of interest, and
then collect all the trials for the final analysis.

 

Hope this is useful.

 

Best,

Jason

 

 

From: eeglablist-bounces at sccn.ucsd.edu
[mailto:eeglablist-bounces at sccn.ucsd.edu] On Behalf Of Mahesh Casiraghi
Sent: Saturday, December 11, 2010 6:34 PM
To: eeglablist at sccn.ucsd.edu
Subject: [Eeglablist] How to correctly break down AR runica() in case of
huge sets.

 

Dear more experienced EEGLabbers and ICA experts,

 

 

supposing one has to work with quite large datsets (several channels, very
high sample rate, long record lengths) and would therefore be unable to load
in memory several gigs of data altogether:

 

A) Is it methodologically problematic to run independent ICAs on subgroups
of trials and then separately perform AR (blinks and scalp detected ECG
components rejection) on each of them?

 

B) Assuming it would not be, as I tend indeed to think, a so recommendable
way, is there a methodologically proof way to combine all the obtained - and
presumably heterogeneous - sphere, weights and weights(-1) matrices in 3
single Sph, W, and W(-1) matrices and then use these new to backproject
after component rejection?

 

C) More precisely, let's suppose we have 700 trials and we run 7 independent
ICAs each time on 100 of them. 

 

a) I would proceed in picking-up separately (subjective criteria, adjust,
faster or whatever one may prefer) the to-be-rejected components,
independently from each subgroup of trials.  

b) I would then remove subgroup by subgroup the respective w(-1) columns and
EEG.icaact rows according to the discarded components.

c) I would merge the obtained 7 EEG.icasphere, the 7 EEG.icaweights, and the
7 EEG.icawinv, in 3 single matrices of equal dimensions, averaging through
nanmean (given the fact we are likely to pick up a different amount of
components from each of the trial subgroups and we would need consistent
matrix dimensions).  

d) I would finally independently backproject subgroup by subgroup using the
same averaged EEG.icawinv and EEG.icasphere and each time the EEG.icaact of
the current subgroup of trials.

 

According to my first speculations, following a->b->c->d we should come up
with something analogous to the output of a big global ICA.

 

Am I wrong?

 

D) Did someone among you already try to run something like that and is
perhaps willing to provide some feedbacks-impressions?

 

 

Cheers,

 

Mahesh

 

 

  

 

Mahesh M. Casiraghi

PhD candidate - Cognitive Sciences

Roberto Dell'Acqua Lab, University of Padova

Pierre Jolicoeur Lab, Univesité de Montréal

mahesh.casiraghi at umontreal.ca

 

I have the conviction that when Physiology will be far enough advanced, the
poet, the philosopher, and the physiologist will all understand each other.

Claude Bernard

 

 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20101216/e24c3a88/attachment.html>