JAX based parallel inference for reactive probabilistic programming - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2022

JAX based parallel inference for reactive probabilistic programming

Résumé

ProbZelus is a synchronous probabilistic language for the design of reactive probabilistic models in interaction with an environment. Reactive inference methods continuously learn distributions over the unobserved parameters of the model from statistical observations. Unfortunately, this inference problem is in general intractable. Monte Carlo inference techniques thus rely on many independent executions to compute accurate approximations. These methods are expensive but can be parallelized. We propose to use JAX to parallelize ProbZelus reactive inference engine. JAX is a recent library to compile Python code which can then be executed on massively parallel architectures such as GPUs or TPUs. In this paper, we describe a new reactive inference engine implemented in JAX and the new associated JAX backend for ProbZelus. We show on existing benchmarks that our new parallel implementation outperforms the original sequential implementation for a high number of particles.
Fichier principal
Vignette du fichier
lctes22.pdf (332.17 Ko) Télécharger le fichier
Origine : Fichiers produits par l'(les) auteur(s)

Dates et versions

hal-03891766 , version 1 (09-12-2022)

Identifiants

Citer

Guillaume Baudart, Louis Mandel, Reyyan Tekin. JAX based parallel inference for reactive probabilistic programming. LCTES'22 - 23rd ACM SIGPLAN/SIGBED International Conference on Languages, Compilers, and Tools for Embedded Systems, Jun 2022, San Diego CA, United States. ⟨10.1145/3519941.3535066⟩. ⟨hal-03891766⟩
41 Consultations
119 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More