Abstract
The aim of probabilistic programming is to automatize every aspect ofprobabilistic inference in arbitrary probabilistic models (programs) so thatthe user can focus her attention on modeling, without dealing with ad-hocinference methods. Gradient based automatic differentiation stochasticvariational inference offers an attractive option as the default method for(differentiable) probabilistic programming as it combines high performance withhigh computational efficiency. However, the performance of any (parametric)variational approach depends on the choice of an appropriate variationalfamily. Here, we introduced a fully automatic method for constructingstructured variational families inspired to the closed-form update in conjugatemodels. These pseudo-conjugate families incorporate the forward pass of theinput probabilistic program and can capture complex statistical dependencies.Pseudo-conjugate families have the same space and time complexity of the inputprobabilistic program and are therefore tractable in a very large class ofmodels. We validate our automatic variational method on a wide range of highdimensional inference problems including deep learning components.