Static Delayed Sampling for Probabilistic Programming Languages
summary
Introduces static delayed sampling that moves delayed-sampling operations to compile time by constructing a programmatic Bayesian network, applying automated marginalization and conditioning on conjugate relationships, and emitting an optimized probabilistic program. Handles features like stochastic indexing and multiplexed outputs conservatively so it avoids full generality but eliminates runtime graph bookkeeping. Implementation targets universal PPLs and generic inference engines, removes dynamic overhead, and demonstrates reduced estimator variance after compile-time transformation.
AI-written summary · about
chapters (5)
- 0:00Start
- 1:49Literature: Other compile time approaches
- 2:20Research Approach
- 3:13From Miking CorePPL to a PBN
- 5:31Conclusion
related talks
17mLazy Knowledge Compilation for Discrete PPLs
13mSemantic Foundations for Laziness in Discrete Probabilistic Programming
19mCompiling Probabilistic Programs for Variable Elimination with Information Flow
19mIncremental Computation for Efficient Programmable Inference in Probabilistic Programs
11mImpulse: Momentously Fast, General, and Portable Probabilistic Programming via(…)
20mVerified Density Compilation for a Probabilistic Programming Language
15mScaling Optimization Over Uncertainty via Compilation
24mAn Intrinsically-typed Probabilistic Programming Language in Coq (Extended Abstr...