Sun Dec 8th through Sat the 14th, 2019 at Vancouver Convention Center
An interesting and potentially influential contribution to the variational inference literature. The authors provides a framework for reinterpreting any Monte Carlo variational objective (e.g. based on IS or particle filtering) as an ELBO with a richer variational distribution, which means that the approach provides a new general way of enriching variational posteriors. The technique is demonstrated on several MC estimators of the likelihood based on variance reduction techniques. The experimental section is currently minimal and is more of a proof-of-concept, but given the significance of the conceptual and theoretical contributions of the paper this is not a significant issue. A conclusion section and the discussion of some missing connections to prior work pointed out by the reviewers, promised by the authors in their response will be very welcome additions to an already strong paper.