Variational Inference with Implicit Distributions

Slides: https://bayesgroup.github.io/bmml_sem/2018/Molchanov_Implicit%20Models_2018_2.pdf Conventional variational inference problems are defined by the likelihood function, the prior distribution and the parametric approximate posterior, which are all usually explicit: we can sample from them, reparameterize them and compute their density. As soon as one component becomes implicit (we can't compute the density), the variational inference becomes intractable. In this talk I will review several approaches that allow us to perform variational inference with implicit distributions. The use of implicit variational inference provides many exciting benefits from fitting an arbitrarily flexible implicit posterior to likelihood-free variational inference. Papers: https://arxiv.org/abs/1702.08235, https://arxiv.org/abs/1511.02386, https://arxiv.org/abs/1808.02078