[Neuroimaging] Interest in modeling library for NiPy

[John Griffiths]

OK well a few navigational points first:


DCM and SPM in general are obviously well developed code bases but leave a
lot to be desired in terms of explicit documentation.


The manual doesn't even begin to touch on half of the stuff squirreled away
inside the 'toolbox' folder

https://github.com/neurodebian/spm12/tree/master/toolbox

A general browse through that folder is well worth doing.



More specifically: many of the main DCM workshorse functions are the ones
in this folder with 'dcm' in the title:

https://github.com/neurodebian/spm12

I think that contains all of the main inversion routines, which don't
necessarily have 'dcm' names, such as

https://github.com/neurodebian/spm12/blob/master/spm_nlsi.m

...as well as high level model comparison functions like

https://github.com/neurodebian/spm12/blob/master/spm_dcm_compare.m




The DCM M/EEG toolbox

https://github.com/neurodebian/spm12/tree/master/toolbox/dcm_meeg

then has lots of model specific things with the (relative to fMRI) more
detailed and more developed M/EEG neurophysiological models.



Also worth looking at the DEM (dynamic expectation maximization) toolbox

https://github.com/neurodebian/spm12/tree/master/toolbox/DEM


and Neural models toolbox

https://github.com/neurodebian/spm12/tree/master/toolbox/Neural_Models



General point: there are a number 'demo' functions littered around; e.g.

https://github.com/neurodebian/spm12/blob/master/toolbox/dcm_meeg/spm_epileptor_demo.m


Which seem to me to be often the best place to look for general
documentation.








On 7 November 2016 at 05:11, Marmaduke Woodman <mmwoodman at gmail.com> wrote:

>
> On Fri, Nov 4, 2016 at 7:18 PM, John Griffiths <j.davidgriffiths at gmail.com
> > wrote:
>
> the distinction is between inference on parameters vs. inference on models
>> (parametric/non-parametric has separate meanings); and not DCM's estimates
>> of effective connectivity parameters per se but rather model
>> evidence/fit/frenergy metrics and comparisons thereof. Certainly it is
>> essential to support both.
>>
>
> I would focus first on the former: an API would allow specification of a
> dataset, a generative model and an inference scheme; the results would be
> inference diagnostics and posteriors.
>
> One could build on that to specific multiple models or a model space and
> comparison criteria.
>
> Anyone with experience in DCM's API might be able to suggest how to make
> that user friendly?
>
>
>> PyMC3 does look like the way to go.
>>
>
> Edward (http://edwardlib.org) is a new one also worth looking at, because
> it builds mainly on TensorFlow. I'm not sure even HMC will scale to full
> size neuroimaging data (though networks with several or tens of nodes would
> work), so it's important to keep the variational schemes available.
>
>
> _______________________________________________
> Neuroimaging mailing list
> Neuroimaging at python.org
> https://mail.python.org/mailman/listinfo/neuroimaging
>
>


-- 

Dr. John Griffiths

Post-Doctoral Research Fellow

Rotman Research Institute, Baycrest

Toronto, Canada

and

Honorary Associate

School of Physics

University of Sydney
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/neuroimaging/attachments/20161107/03b7fe84/attachment.html>