Black-Box Variational Inference for Probabilistic Programs
summary
Claims that black‑box variational inference applies to universal probabilistic programs by converting the program into a parameterized approximating program: replace parameters with θ and drop observe statements, run that program to produce traces, compute prior and likelihood weights, and use Monte Carlo gradient estimates of the KL objective to update θ by stochastic gradient descent. Proposes a lower‑variance estimator exploiting the program’s Markov‑chain factorization, but reports no implementation yet and no empirical results.
AI-written summary · about
chapters (7)
- 0:00Intro
- 1:46Inference engine for probabilistic programs
- 13:16Importance sampling
- 17:07Variational approximation
- 20:17Minimising KL divergence
- 22:31Key observation
- 25:26Black-box variational inference
related talks
21mProbabilistic Programming with Programmable Variational Inference
31mProbabilistic Programming: The What, Why and How
37mVariational Inference in Python
06mPitfalls of Full Bayesian Inference in Universal Probabilistic Programming
21mGuaranteed inference for probabilistic programs: a parallelisable, small-step o...
24mExact Bayesian Inference for Loopy Probabilistic Programs Using Generating Functions
46mBlueprints for a Universal Reasoning Machine
01h 07mProbabilistic Programming and Readable Models