Recursive Learning of Asymptotic Variational Objectives