Dan Simpson • 5/7/2024

An unexpected detour into partially symbolic, sparsity-expoiting autodiff; or Lord won’t you buy me a Laplace approximation

This article details a deep dive into implementing Laplace approximations, a method for approximating distributions, using the JAX library. It covers the mathematical theory, then focuses on practical implementation challenges involving sparse automatic differentiation and manipulating JAX's intermediate representation (jaxpr) to optimize gradient computation for a logistic regression example, aiming for performance close to a hand-coded gradient.

0 comments

#Jax #Automatic Differentiation #Laplace Approximation