Iterative Refinement of the Approximate Posterior for Directed Belief Networks