another motivation for MML or VB

[Andrew Gelman]

Hi all. We had an exchange on stan-users regarding the slowness of models with latent multivariate random variables. This made me think that this could be a good test example for MML and VB methods that can integrate over these latent variables. Worth thinking about as we develop and implement these approximate algorithms.
A

[Michael Betancourt]

Isn’t this going to be exactly the kind of model with high enough curvature
that MML and VB would yield poor approximations? I hate to be the wet
blanket, but the same pathologies that make Euclidean HMC slow also
make these approximations slow, so there’s only only a narrow slice in
model space where MML and VB could flourish.

On Dec 1, 2015, at 7:02 AM, Andrew Gelman <gel...@stat.columbia.edu> wrote:

> Hi all. We had an exchange on stan-users regarding the slowness of models with latent multivariate random variables. This made me think that this could be a good test example for MML and VB methods that can integrate over these latent variables. Worth thinking about as we develop and implement these approximate algorithms.
> A
>

> --
> You received this message because you are subscribed to the Google Groups "stan development mailing list" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to stan-dev+u...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.