Advice on using variational Bayes

Rationale of using VB

in building models for third party users the execution time is an essential point of success. It is a pivotal point for most users between using of not. (I am already implementing models with multi-threading/MPI)

The hope

That building an incredibly well behaved model will allow the use of vb

The question

Although I could find sporadic information, I failed to find a guideline. What are the requirements for a successful application of VB? For example:

  • All parameters should have normal shape posterior? (meaning that if a parameter is a proportion, and could be asymmetric I have to forget about VB)
  • All parameters have to be unitary? (can parameter be in the range from 0.1-1 or 1-10?)
  • … other important issues to consider before attempting the use of VB?

Thanks!

I don’t think we have any guidelines. ADVI has failed for well conditioned linear regression with fixed parameters. Sorry.