[Neuroimaging] Interest in modeling library for NiPy
j.davidgriffiths at gmail.com
Fri Nov 4 14:18:07 EDT 2016
On 4 November 2016 at 13:47, Marmaduke Woodman <mmwoodman at gmail.com> wrote:
> > How can we help?
> It would be really really helpful to get feedback on the scope: what sort
> of functionality would you want from such a library? If you've ever done
> modeling work, what tripped you up or made life difficult?
> I see a few main dimensions of scope & functionality:
> First, in the projects I've seen using TVB and DCM, I see network models
> used both parametrically and non parametrically. By parametrically, I mean,
> for example, that we do a parameter sweep over coupling strength, and
> compare empirical and simulated FC, and then look at, for example, a group
> wise difference of the best coupling strength. By non parametrically, I
> have in mind things like DCM's estimates of effective connectivity, where
> many parameters are being estimated and they are interpreted as an
> ensemble, not individually. This is naturally a slippery slope and some
> modeling questions lie between the two extremes.
I think the distinction you're pointing to here 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.
> The second dimension of scope is in terms of the methods and numerics
> implemented. Simple time-stepping schemes for differential equations are
> easy to implement, but making them high-performance is less so (think
> CUDA/OpenCL). Bayesian inversion is really neat, but requires computing
> gradients or using packages like PyMC3 or Stan.
Definitely agree that model inversion/fitting should be a priority design
consideration from the very start. PyMC3 does look like the way to go.
> Finally, I would assume we're mainly interested in human or primate
> neuroimaging, so modalities like fMRI & MEG, and maybe invasive clinical
> modalities too. As more of a methods library, this would be a detail I
> guess, and I would expect to delegate I/O, formatting, etc to the
> respective libraries.
Incidentally: I understand that there is a new MNE-NEURON project on the go
(PIs Matti Hamalainen & Stephanie Jones), that will be looking to fit
lower-level (compartmental) neuron models to MEG signals from humans and
animals. Could well be a lot of overlap on model fitting problems.
> Again, consider this a RFC and let me know what you think.
> Neuroimaging mailing list
> Neuroimaging at python.org
Dr. John Griffiths
Post-Doctoral Research Fellow
Rotman Research Institute, Baycrest
School of Physics
University of Sydney
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Neuroimaging