[Eeglablist] Comments regarding talk

Mon Aug 12 19:11:48 PDT 2019

Dear Makoto,

Thank you for your answers and the resources you linked to - they have been very clarifying. I appreciate your attention.

Guillermo

Obtener Outlook para Android<https://aka.ms/ghei36>

On Mon, Aug 12, 2019 at 9:49 PM -0400, "Makoto Miyakoshi" <mmiyakoshi at ucsd.edu<mailto:mmiyakoshi at ucsd.edu>> wrote:

Dear Guillermo,

Thank you for your further comment.

> Do you foresee any kind of issue that might limit the applicability of
such framework?

To confirm what you mean by 'framework' here, it refers to EEG's hard
problem i.e., 1) source patch size should be > 1 inch^2 to be detected at
scalp electrodes (i.e., rounding down all the active sources with < 1
inch^2); 2) gyral sources are preferably detected; 3) the targeted
activation must occur in neocortex (a few exceptions mentioned below). If
this 'framework' does not apply... for example, auditory brain stem
response (ABR) seems to be a valid exception, since it is localized in
brain stem but massive averaging gains huge SNR to make it
scalp-measurable. Actually, > 1inch^2 rule refers to the condition for
NON-AVERAGED, 20-30 uV at the scalp measurement. If somatosensory evoked
potential is averaged for 500 times, maybe the source regions could be much
smaller than 1 inch^2.

 > Besides what you mention, I was also referring to the assumptions that
ICA must hold in order to be consistent. My concern specifically focus on
the election of ICA algorithm, as there are many available (each with its
own particularities), results after application of ICA can be different
when applying, for example, an algorithm that relies on Mutual Information
in relation to when we apply an algorithm based on Tensor Diagonalization.

Well, let's not extend this discussion to tensor decomposition. It depends
on whether you see the tensor decomposition as a 'natural' extension of the
matrix decomposition, and unfortunately I am not knowledgeable enough to
judge it (for example, it seems hard to compare performance between 2D vs.
3D decompositions, it seems to me apple vs. orange.) So let's assume we
stick to matrix factorization for the same of argument.

> This uncertainty is what also encouraged me to think of it as part of the
pessimism. However, perhaps I'm exceeding a bit and this can be addressed
successfully by other steps.

Are you saying that the above argument also depends on the variations of
ICA (including its tensor extensions)? But their assumptions should be more
or less the same (I'm quite naively saying it.) 'ICA must hold' is
admittedly a critical assumption, but it seems to me where ICA's
assumptions fails, we cannot expect ordinary EEG data.

SCCN's best reasoning about ICA has been that 'Reducing mutual information
increased dipolarity of the ICs' (Delorme et al., 2012). Once I named this
notion as independence-dipolarity identity (IDID). If I rephrase it (with
my colleague Hiro Tanaka's words), *solving temporal problem also solves
spatial problem is a non-trivial property of ICA. *This has been the Holy
Grail for the ICA users. This is indeed fascinating to confirm, which is
why ICA is a working solution for me. But recently, I came to realize that
invalidly deep dipoles with valid time domain data, which you often see in
the final results, may be associated with invalidly large dipole moment.
What if the ICs with 'highly dipolar' scalp topos actually have
nonsensically large dipole moments? Does IDID claim still hold? This must
comes from the conceptual contradiction that dipole = 'point', while ICA
could generate scalp projections from 'spatially distributed sources' which
dipole (as a point) cannot represent. So after all, this could be simply
violation of assumption of a dipole model.

Makoto

On Mon, Aug 12, 2019 at 8:55 PM GUILLERMO SAHONERO ALVAREZ <
guillermo.sahonero at ucb.edu.bo> wrote:

> Dear Makoto,
>
> Keeping up the conversation:
>
> The whole hypothesis perhaps points out to something like a limitation
>
>
> principle. What I try to mean is that, perhaps the amount of information is
> limited by definition - not only from a practical approach but from a
> theoretical framework. Did you try to address in any way this?
>
> Yes exactly..
>
>   *   Do you foresee any kind of issue that might limit the applicability
> of such framework?
>
>
>
>
> As ICA algorithms suppose working principles, the 10-20 brain ICs you
>
>
> referred would not be guaranteed to be the same in every subject. This
> might also be a component to add into the "pessimism".
>
> I like you say 'working principles'. Yes, even after critical thinking, I
> still like to use ICA as a working principle.
>
> In scalp EEG study, there are some factors that are neuroscientifically
> non-interesting but still determine what can be measured at the scalp
> level, such as individual differences in head geometry, cortical
> gyrification patterns, total skull conductivity, etc. For example, if a
> cortical region of interest in one subject happened to be centrally located
> in a sulcus rather than a gyrus, this alone makes scalp-level measurement
> harder. Personally I have never seen a ratio of gyral area vs. sulcal area,
> but such a statistics would help us to estimate how much of our cortical
> activity is obscured just because of this unfortunate fact.
>
> I do trust ICA's capability to remove mutual information to recover
> temporal independence (given the stationarity assumption holds--which is
> never be true in reality), but if our 'raw signal' is compromised as I
> discussed, we can't expect that ICA 'recovers' information, but it is
> rather projecting blurred scalp projection back into the brain; it seems to
> work as a deblurr filter, but it may not 'recover' the lost information. It
> is a pessimistic situation, but not independent one but it is already
> included by EEG's hard problem I proposed.
>
>   *   Besides what you mention, I was also referring to the assumptions
> that ICA must hold in order to be consistent. My concern specifically focus
> on the election of ICA algorithm, as there are many available (each with
> its own particularities), results after application of ICA can be different
> when applying, for example, an algorithm that relies on Mutual Information
> in relation to when we apply an algorithm based on Tensor Diagonalization.
> This uncertainty is what also encouraged me to think of it as part of the
> pessimism. However, perhaps I'm exceeding a bit and this can be addressed
> successfully by other steps.
>
>
>
>
> Actually, I would like to know if you have developed some short studies
>
>
> that allowed to observe some kind of pattern of the ICs that are linked to
> the effective degrees of freedom - maybe even the most important EEG
> channels.
>
> No, I have never thought of it. If we agree to use single-model ICA as a
> working principle, there could be some approach. It seems possible to draw
> an empirical conclusion that some group of scalp channels receives central
> projection by ICs more frequently than other channels... but we have to
> consider the dipole orientation in doing this, which could be another
> neuroscientifically non-interesting factor.
>
>   *   I agree entirely. But, I think it might be useful to know such group
> of scalp channels. This way, even computational models can focus more on
> data that comes from specific channels and perhaps increase the amount of
> useful information that is used in EEG based technologies - like BCIs.
>
>
>
>
> ECoG relies on having very specific local arrays of microelectrodes. In
>
>
> my opinion, this might not be very suitable to represent the entire brain
> complexity.
>
> The ideal ECoG recording is to record from ALL of the cortical surface. If
> it is difficult (and it is difficult), I think we should try to minimize
> the inter-electrode gap. In ECoG, volume conduction gets attenuated very
> rapidly. But if the electrode is touching most of the cortical surface, we
> can record multiple small source activity as a linear mixture.
>
>
>
> On the other hand, a combination of EEG with ECoG could represent an
>
>
> increase of the amount of EDoF, do you know of any work that addressed such
> idea?
>
> There are some scalpEEG-ECoG simultaneous recording papers and data
> available. Neurotycho is one such database available
> http://neurotycho.org/ The
> problem of scalpEEG-ECoG data is that the data is rare because when ECoG is
> available recording scalpEEG seems almost unnecessary
>
> By the way, when I returned from SFI I saw this news. I welcome these
> ambitious approaches. https://www.biorxiv.org/content/10.1101/703801v2
> Repeating
> one-bit information generation by repeating experiments is a traditional
> scientific model of clarifying something, and scalp-recorded EEG may be ok
> for that purpose. But these silicon-valley guys seems to have more direct
> and ambitious ideas in mind, and they have budget to test the ideas too.
>
>   *   Thanks for sharing. I will check those works.
>
>
>
> Again, thank you very much for your interest.
>
> Thanks for answering.
>
> Guillermo
>
>
>
> Makoto
>
> On Tue, Jul 30, 2019 at 6:04 PM GUILLERMO SAHONERO ALVAREZ <
> guillermo.sahonero at ucb.edu.bo>
> wrote:
>
>
>
> Dear Makoto,
>
> I would like to share some comments about your one hour talk at Santa Fe
> Institute:
>
>   *   The whole hypothesis perhaps points out to something like a
> limitation principle. What I try to mean is that, perhaps the amount of
> information is limited by definition - not only from a practical approach
> but from a theoretical framework. Did you try to address in any way this?
>   *   As ICA algorithms suppose working principles, the 10-20 brain ICs
> you referred would not be guaranteed to be the same in every subject. This
> might also be a component to add into the "pessimism". Actually, I would
> like to know if you have developed some short studies that allowed to
> observe some kind of pattern of the ICs that are linked to the effective
> degrees of freedom - maybe even the most important EEG channels.
>   *   ECoG relies on having very specific local arrays of microelectrodes.
> In my opinion, this might not be very suitable to represent the entire
> brain complexity. On the other hand, a combination of EEG with ECoG could
> represent an increase of the amount of EDoF, do you know of any work that
> addressed such idea?
>
> I hope we could exchange some ideas. Thank you for your attention.
>
> Guillermo
>
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to
> eeglablist-unsubscribe at sccn.ucsd.edu 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
>
>
>
>
>
>
> _______________________________________________
> 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
Assistant Project Scientist, Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
_______________________________________________
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