Oops I Took A Gradient: Scalable Sampling for Discrete Distributions

     Will Grathwohl 1 2 Kevin Swersky 2 Milad Hashemi 2 David Duvenaud 1 2 Chris J. Maddison 1

             Abstract
   We propose a general and scalable approximate
   sampling strategy for probabilistic models with
   discrete variables. Our approach uses gradients
   of the likelihood function with respect to its dis-
   crete inputs to propose updates in a Metropolis-
   Hastings sampler. We show empirically that this
    ...