Title: Monte Carlo and variational methods: Bridging the gap
Authors: Christian Andersson Naesseth - University of Amsterdam (Netherlands) [presenting]
Abstract: The goal of inference is to reach conclusions based on evidence and reasoning; to use data and models to find patterns and answer questions. Practical problems often result in situations where exact inference is intractable and we must resort to approximations. The two main paradigms for approximate inference are sampling-based methods, also known as Monte Carlo methods, and optimisation-based methods, also known as variational methods. We will explore the interplay between Monte Carlo and variational methods, focusing on recent work combining MCMC and variational inference using the forward Kullback-Leibler divergence.