A0800
Title: Approximate sampling and estimation of partition functions using neural networks
Authors: George Cantwell - University of Cambridge (United Kingdom) [presenting]
Abstract: The closely related problems of sampling from a distribution known up are considered to be a normalizing constant and estimating said normalizing constant. The purpose is to show how variational autoencoders (VAEs) can be applied to this task. VAEs are trained to fit data drawn from an intractable distribution in their standard applications. The logic and train of the VAE are inverted to fit a simple and tractable distribution on the assumption of a complex and intractable latent distribution specified up to normalization. This procedure constructs approximations without the use of training data or Markov chain Monte Carlo sampling. The method on three examples are illustrated: the Ising model, graph clustering, and ranking.
