Automatic Differentiation of Programs with Discrete Randomness

Abstract: Automatic differentiation (AD), a technique for constructing new programs which compute the derivative of an original program, has become ubiquitous throughout scientific computing and deep learning due to the improved performance afforded by gradient-based optimization. However, AD systems have been restricted to the subset of programs that have a continuous dependence on parameters. Programs that have discrete stochastic behaviors governed by distribution parameters, such as flipping a coin with probability p of being heads, pose a challenge to these systems because the connection between the result (heads vs tails) and the parameters (p) is fundamentally discrete.

This talk discusses recent work on tackling this problem to develop low-variance, unbiased estimators, and their application to discrete stochastic simulations in the sciences.

Bio: Gaurav Arya is an undergraduate at MIT majoring in mathematics and computer science. He has been doing research in the Nanostructures and Computation Group at the Julia Lab.