The Gumbel-Softmax is a continuous distribution over the simplex that is often used as a relaxation of discrete distributions. Because it can be readily interpreted and easily reparameterized, the Gumbel-Softmax enjoys widespread use. We show that this relaxation experiences two shortcomings that affect its performance, namely: numerical instability caused by its temperature hyperparameter and noisy KL estimates. The first requires the temperature values to be set too high, creating a poor correspondence between continuous components and their respective discrete complements. The second, which is of fundamental importance to variational autoencoders, severely hurts performance. We propose a flexible and reparameterizable family of distributions that circumvents these issues by transforming Gaussian noise into one-hot approximations through an invertible function. Our construction improves numerical stability, and outperforms the Gumbel-Softmax in a variety of experiments while generating samples that are closer to their discrete counterparts and achieving lower-variance gradients. Furthermore, with a careful choice of the invertible function we extend the reparameterization trick to distributions with countably infinite support.
Speakers: Andres Potapczynski, Gabriel Loaiza-Ganem, John Cunningham