GradInf: Gradient Estimation as Probabilistic Inference
summary
Introduces GradInf, a source-to-source system that reduces gradient-of-expectation problems to probabilistic inference by synthesizing an inference target via coupling and factorization, then applies standard inference (importance sampling, SMC, exact elimination) plus automatic differentiation. Uses an information-flow type system and quasi‑Borel semantics to prove soundness for higher‑order programs. Expresses prior estimators (pathwise, score, phantom) and in a queue case swaps to exact elimination to obtain 100× variance reduction; artifacts and code are available.
AI-written summary · about
chapters (9)
- 0:00Gradient estimation intro
- 2:50Gradient inference approach
- 7:05Coupling for variance reduction
- 9:52Factorization for variance reduction
- 12:19Probabilistic inference role
- 13:44Difference to gradient conversion
- 14:47Gradient inference applications
- 15:07Contributions and future work
- 15:30Q&A
related talks
21mProbabilistic Programming with Programmable Variational Inference
11mExact Inference for Discrete Probabilistic Programs via Generating Functions
31mProbabilistic Programming: The What, Why and How
14mExact Probabilistic Inference Using Generating Functions
26mADEV: Sound Automatic Differentiation of Expected Values of Probabilistic Progra...
18mA Language and Smoothed Semantics for Convergent Stochastic Gradient Descent
28mSmoothness Analysis for Probabilistic Programs with Application to Optimised Var...
45mPolicy Gradient: Optimal Estimation, Convergence, and Generalization beyond Cumulative Rewards