Sample from categorical distribution

It is possible to (approximately) sample from a categorical distribution in a continuous, differentiable form. This ICLR 2017 paper introduces the Gumbel-Softmax distribution which relies on the Gumbel distribution to sample one-hot vectors from a categorical distribution.

The blog post I linked shows how the Gumbel-Max trick is equivalent to a softmax output. And the paper shows that you can sample.

This Gumbel-Softmax distribution can be used for neural architecture search, like for instance the Differentiable Neural Architecture Search (DNAS).