[Eeglablist] Component clustering and PCA reduction

[Makoto Miyakoshi]

Dear Clemens,

This clarification may be of interest for other subscribers too, so let me
get back to the list to continue.

Sorry for slow response. I've been swamped by a project I've been working
on since January, which is finished (I believe) yesterday.

> Did I understand correctly, that, to take ERSPs as an example, the data,
wich is hundreds of time/frequency points is reduced to principal
components that explain the most variance? In components that for example
capture a lot of fm-theta, would this (first) principal component take its
weight from the time/frequency points in 4-8 Hz range, high in power, and
some latency? In other words, there a hundreds of dimensions, if I reduce
to five dimensions, I am left with, mostly theta, some alpha, some noise,
etc.?

This is my first time to check this process in code so I could be wrong.
For the case of ERSP/ITC, std_preclust() line 362 says X, which should be
frequency x time, is reshaped to 1 x numel(X), which means time-frequency
is vectorized. For example, suppose you have ERSP/ITC matrix of 30 (freqs)
x 200 (times) x 50 (ICs), you'll run runpca() at line 404 of the same
function on 6000 x 50 matrix. If you specify ERSP/ITC dimension of 10, this
should be 10 x 50 matrix after dimension reduction. So the vectorized
time-frequency data are dimension-reduced.

> I am sorry to bother you, perhaps you can also suggest some literature
where people have done similar things. (I have read Julie Ontons paper from
2005 on FMT cluster, Still I am not so confident if I get the PCA thing)

I have never seen a publication detailing this part. However, this way of
using PCA to reduce data dimension is extremely popular in the engineering
field, and no need to validate... that's how I feel.

Makoto



On Thu, Jun 9, 2016 at 2:13 AM, Clemens DICKHUT <clemens.dickhut at uni.lu>
wrote:

> Dear Makoto,
>
> Thanks a lot for your reply. I have read the online manual thoroughly,
> however, the part where it is described how PCA reduces the dimensions of
> the respective measures is a little vague… at least I am not able to grasp
> it.
>
> Did I understand correctly, that, to take ERSPs as an example, the data,
> wich is hundreds of time/frequency points is reduced to principal
> components that explain the most variance? In components that for example
> capture a lot of fm-theta, would this (first) principal component take its
> weight from the time/frequency points in 4-8 Hz range, high in power, and
> some latency? In other words, there a hundreds of dimensions, if I reduce
> to five dimensions, I am left with, mostly theta, some alpha, some noise,
> etc.?
>
> I am sorry to bother you, but this step really is a black box to me…
>
> I am sorry to bother you, perhaps you can also suggest some literature
> where people have done similar things. (I have read Julie Ontons paper from
> 2005 on FMT cluster, Still I am not so confident if I get the PCA thing)
>
> All the best,
> Clemens
>
> __________________________________
> Clemens Dickhut, M.Sc., (PhD Student)
> Institute for Health and Behaviour
> Research Unit INSIDE
> University of Luxembourg - Campus Belval
> Maison des Sciences Humaines
> 11, Porte des Sciences, R. 04 415
> L-4366 Esch-sur-Alzette
> Tel.: (+352) 46 66 44 9536
>
>
>
> On 07 Jun 2016, at 01:26, Makoto Miyakoshi <mmiyakoshi at ucsd.edu> wrote:
>
> Dear Clemens,
>
> > As for example, I want to reduce the dimensions of ERSP/dipoles,
> spectra, given all ICs, which dimensions, what are the dimensions that
> would be reduced for the respective measures?
>
> If I remember correctly, EEGLAB manual says
>
> 1. Set the parameters so that total number of dimensions across all the
> measures are <20-30 because k-means clustering performance becomes worse if
> >30.
>
> 2. Do not use 'final dimension' thing.
>
> I think it's a good idea to keep it below 20 to see how it works.
> I personally recommend you use just dipole and just a little bit (weight 1
> - 3) of spectrum with dimension 10-15. It's a good idea to avoid the
> measure which you'll test using statistics in the end to avoid 'double
> dipping' issue (if you don't know what it is, google 'voodoo correlation'.)
>
> Makoto
>
>
> On Tue, May 24, 2016 at 1:42 AM, Clemens DICKHUT <clemens.dickhut at uni.lu>
> wrote:
>
>> Dear all,
>>
>> I have some trouble understanding the reduction of dimensionality on
>> measures selected for component clustering. (STUDY —> build pre-clustering
>> array)
>>
>> As for example, I want to reduce the dimensions of ERSP/dipoles, spectra,
>> given all ICs, which dimensions, what are the dimensions that would be
>> reduced for the respective measures?
>>
>> I am grateful for any input.
>>
>> Best,
>> Clemens
>>
>> __________________________________
>> Clemens Dickhut, M.Sc., (PhD Student)
>> Institute for Health and Behaviour
>> Research Unit INSIDE
>> University of Luxembourg - Campus Belval
>> Maison des Sciences Humaines
>> 11, Porte des Sciences, R. 04 415
>> L-4366 Esch-sur-Alzette
>> Tel.: (+352) 46 66 44 9536
>>
>>
>>
>>
>> _______________________________________________
>> 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/20160701/9bbe7e86/attachment.html>