REBAR: Low-variance, unbiased gradient estimates for discrete latent variable models

   Abstract