I created a fairly complex model using Edward. It does run amazingly fast. However it is also true that at the end some parameters come out spot on while the smaller values got inflated. Understand that inaccuracies is part of the characteristic that come which VI. I am wondering though what kind of applications would be the sweet spot for VI which can tolerate some inaccurately inferred parameters ?
On this front, I was pointed to this paper on Pareto Smoothed Importance Sampling
as a way to measuring whether an VI result is trustworthy, and make adjustments to improve the RMSE. Is there any plan to incorporate this into Edward?