Guaranteed inference for probabilistic programs: a parallelisable, small-step o...
summary
Presents a small-step operational semantics for probabilistic programs using a product of Markov kernels, handling mixed discrete/continuous distributions, conditioning and unbounded loops. Derives an exact sampling algorithm that vectorizes kernel applications and is SIMD-parallelized, turning sequential Monte Carlo into batch tensor operations. Proves a finite-approximation theorem that gives computable lower bounds on observational probabilities and shows a TensorFlow proof-of-concept that yields competitive accuracy and runtimes versus state-of-the-art tools.
AI-written summary · about
chapters (9)
- 0:00Probabilistic programs intro
- 2:09Explicit distributions in code
- 4:29Static vs dynamic inference
- 7:32Formalizing probabilistic programs
- 10:42Observational semantics
- 13:18Vectorized Monte Carlo
- 15:58Correctness and bounds
- 17:50Approximation bounds
- 19:18Q&A
related talks
01h 03mA Personal Viewpoint on Probabilistic Programming
31mProbabilistic Programming: The What, Why and How
19mIncremental Computation for Efficient Programmable Inference in Probabilistic Programs
24mExact Bayesian Inference for Loopy Probabilistic Programs Using Generating Functions
26mSemi-symbolic Inference for Efficient Streaming Probabilistic Programming
20mProbabilistic Programming with Stochastic Probabilities
32mProbabilistic and Differentiable Programming in Scientific Simulators
47mThe State of the Art for Probabilistic Programming