GradInf: Gradient Estimation as Probabilistic Inference

DOI10.1145/3808321MaRDI QIDQ7284003arXivFDO

Authors Gaurav Arya, Mathieu Huot, Moritz Schauer, Alexander K. Lew, Feras A. Saad

Publication date 8 July 2026

Published in arXiv

Tldr We introduce a novel approach called gradient inference which reformulates gradient estimation as probabilistic inference, allowing us to leverage probabilistic programming tools to design efficient and sound gradient estimators for complex probabilistic programs. We develop GradInf, a probabilistic programming system that automates gradient inference with provable correctness guarantees, enabling the construction of state-of-the-art gradient estimators.

Full work available at URL https://doi.org/10.1145/3808321

arXiv classification

cs.LGcs.PL

Cites work

• An efficient finite difference method for parameter sensitivities of continuous time Markov chains
• An introduction to sequential Monte Carlo
• Bisimulation through probabilistic testing
• Correctness of automatic differentiation via diffeologies and categorical gluing
• Coupling proofs are probabilistic product programs
• Efficiency of the Girsanov transformation approach for parametric sensitivity analysis of stochastic chemical kinetics
• Fast and correct gradient-based optimisation for probabilistic programming via smoothing
• Logical relations and the typed λ-calculus
• On the versatility of open logical relations. Continuity, automatic differentiation, and a containment theorem
• Perturbation analysis via coupling
• Relational reasoning for Markov chains in a probabilistic guarded lambda calculus
• Relational reasoning via probabilistic coupling
• Simple statistical gradient-following algorithms for connectionist reinforcement learning
• Smoothed (conditional) perturbation analysis of discrete event dynamical systems
• Stochastic lambda calculus and monads of probability distributions
• The art of differentiating computer programs. An introduction to algorithmic differentiation.

This page was built for publication: GradInf: Gradient Estimation as Probabilistic Inference

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q7284003)