The ability to backpropagate stochastic gradients through continuous latent
distributions has been crucial to the emergence of variational autoencoders and
stochastic gradient variational Bayes. The key ingredient is an unbiased and
low-variance way of estimating gradients with respect