I have been meaning to write this post about the Gumbel-softmax distribution for several months, but I put it on a back burner after I had dug myself into a hole of deep confusion and couldn’t get out. After some encouragement from my advisor, I decided to pick it up again, and this time I think I was able to figure things out.*1 So in this post, we are going to learn how the Gumbel-softmax distribution can be used to incorporate categorical distributions into algorithms that use neural networks...