Data-oriented Design for Differentiable, Probabilistic Programming

This remote video presentation from the LAFI 2025 workshop explores a reimplementation of LazyPPL's core data structure using data-oriented design techniques. Learn how researchers Owen Lynch, Maria-Nicoleta Craciun, Younesse Kaddar, and Sam Staton from the University of Oxford optimized the lazy rose tree structure to reduce memory usage and enhance performance. The presentation details their implementation of a specialized reverse-mode autodiff graph for computing gradients of lazy rose tree values with respect to probabilistic program scores. This 20-minute talk was presented at the Languages for Inference (LAFI) workshop on January 19, 2025, as part of the ACM SIGPLAN-sponsored POPL 2025 conference.