The Taub Faculty of Computer Science Events and Talks

Reparameterization of variational auto-encoders with continuous random variables is an effective method for reducing the variance of their gradient estimates. Our work optimizes the discrete VAE objective directly, using its Gumbel-Max reparameterization. This optimization technique propagates gradients through the reparameterized argmax, and replaces MCMC sampling by the difference of gradients of two argmax predictions. This realization provides the means to learn latent representations in cases when evaluating the argmax operation is tractable while evaluating the softmax operation is intractable.