When reading Hoffman’s paper on stochastic variational inference, it is said that mean field VI and the associated factorization of individual variables is a root strategy for scalable inference, which implies loss of accuracy in individual parameter approximation in favor of global distribution approximation. When I read the sample models in Edward, it does not appear that factorization is used at all. is mean field VI somehow implied when calling ed.VariationalInference? Or is it something you explicitly build into your model?
It’s something you explicitly build into your variational approximation.