Max Hird
University of Waterloo
Room: M3 3127
Resolving the Incongruity between Markov chain Monte Carlo and Variational Inference with the Occlusion Process
Variational inference (VI) can be very fast, although it is inexact. Markov chain Monte Carlo (MCMC) is asymptotically exact but slow due to the fact that it is inherently serial. It is therefore an attractive prospect to combine the two to create an inference algorithm that is both fast and asymptotically exact. In doing so one encounters a fundamental tension between the two methods: VI produces an object (the variational approximation) that is fundamentally global whereas general purpose MCMC algorithms are fundamentally local. Naïvely combining the two may result in poor performance. We propose the Occlusion Process to resolve this incongruity and combine the best aspects of either method. Joint work with Florian Maire, available to view on arXiv.