Adversarial sequential Monte Carlo

How can we perform efficient inference in directed probabilistic models with intractable posterior distributions? We introduce a new technique for improving finite-sample inference approximations by learning highly flexible proposal distributions for sequential importance samplers, such as particle filters. We represent proposal distributions as implicit generative models, that is models that you can sample from but which do not have an explicit parametric form, and train them using variational inference rephrased as a two-player game, hence establishing a principled connection between Sequential Monte Carlo (SMC) and Generative Adversarial Networks (GANs). Our approach achieves state-of-the-art performance on a range of inference problems.