[FieldTrip] Set regularization parameter after SSS
matti.stenroos at aalto.fi
Tue Aug 21 15:16:23 CEST 2018
Hi Vladimir, Robert,
Thanks for your input!
I've been playing with linear estimates & SSS-processed data. Based on
my experience, I agree with Vladimir that dropping the rank to 60--70
(depending on the number of max filter coeffs etc) is the way to go. If
whitening (via noise covariance matrix) is used, some regularization to
the whitener is probably appropriate.
I don't have strong opinion regarding "only gradiometers" or
"gradio+magneto". I'd say there is no general answer; rather it depends
on what kind on inverse method you apply = how you use your data there.
In resolution analysis of minimum-norm estimates, the difference is small.
Data fusing = sensor scaling should not be that difficult; pre-weighting
the data with 1/sqrt(mvars), where mvars has mean noise/data variances
per sensotype. But, as the Max reconstructs all data using the same
multipole series with ~64 components, rank will not really increase ---
if there has been 64 max filter components, useful ranks of
magnetometers, gradiometers, and the whole MEG are of the order of 60--64.
On 2018-08-21 15:02, Vladimir Litvak wrote:
> Hi Luca and Matti,
> We have recently looked into beamforming on Elekta with several
> methods people including Robert and Alex Gramfort. One thing we
> discovered is that beamforming can completely fail when small
> regularisation (like 1% or 5% which most people use) is applied to
> data after SSS. The theoretical underpinning for this has not been
> figured out yet, but in practice, the more robust thing to do is to
> reduce the data dimensionality and set lambda to zero. I attach an
> example developed for the phantom data we got from Aston MEG that
> shows how this can be done in FT. If you want, I can share the phantom
> example with you so that you can verify that this approach works.
> Regarding sensor type fusion, I would suggest that you only limit your
> analysis to one sensor type. Most Neuromag people use the planar
> gradiometers. The reason for this suggestion is that after SSS,
> channels of both types are linear combinations of the same basis
> vectors and contain redundant information so trying to fuse them is
> more a headache than anything else. There is a paper saying this in
> more detail http://www.mdpi.com/1424-8220/17/12/2926/html
> On Mon, Aug 20, 2018 at 2:35 PM Luca Kaiser <luca.kaiser at web.de> wrote:
>> Hi Matti,
>> sure-sorry and thanks for your quick reply. So here is what I am doing using the average covariance matrix (so data covariance).
>> avg_data=ft_timelockanalysis(cfg, data);
>> cfg.lcmv.fixedori= 'yes';
>> cfg.lcmv.lambda= '10%'; %0% rank deficient data-use stronger regularization??
>> lcmv_avg=ft_sourceanalysis(cfg, avg_data);
>> Gesendet: Montag, 20. August 2018 um 15:10 Uhr
>> Von: "Matti Stenroos" <matti.stenroos at aalto.fi>
>> An: "Luca Kaiser" <luca.kaiser at web.de>
>> Betreff: Re: [FieldTrip] Set regularization parameter after SSS
>> Dear Luca,
>> I think you'd need to tell a bit more, as the meaning of "lambda"
>> depends on the algorithm you are using. Also "the covariance matrix" is
>> not uniquely defined --- are you talking about noise covariance or
>> measurement covariance?
>> In general, the eigenvalue idea you had read does not make sense --- the
>> eigenvalue text deals with condition numbers, while lambda does, in
>> general, not relate directly to that. The advice might have been that
>> "set lambda so that the ratio of largest and smallest eigenvalues is
>> also working with low-rankingMax-filtered MEG and coincidentally
>> having a lot of covariance metrics on screen at the moment
>> On 2018-08-20 15:39, Luca Kaiser wrote:
>>> Dear FieldTrip community,
>>> I am using ft_sourceanalysis on SSS preprocessed (neuromag) data. I
>>> wonder if there are any suggestions on how to set lambda in this case?
>>> My covariance matrix is rank deficient (rank 60, 306 sensors). I read in
>>> the mailing list that I would need to look at the eigenvalues of XX' and
>>> set lambda to be 1/1000 of the largest eigenvalue. However, I do not
>>> really understand why to divide by 1/1000?
>>> Every help would be very much appreciated,
>>> fieldtrip mailing list
>> fieldtrip mailing list
>> fieldtrip mailing list
More information about the fieldtrip