Structured variational inference

Is there any way to specify dependencies between variational parameters, or perform structured variational inference (http://proceedings.mlr.press/v38/hoffman15.pdf)?

There’s a structured variational approximation in the sigmoid_belief_network.py example. I recall fixing a bug to get it to work so you may need Edward’s dev version.

Matt’s SSVI isn’t implemented but an example building and applying it are welcome of course.