Semi-symbolic Inference for Efficient Streaming Probabilistic Programming
summary
Semi-symbolic inference implements an automatic Rao-Blackwellized particle filter in ProbZelus by manipulating symbolic expression trees, using a central swap operation to reorder dependencies, replacing observed roots with Delta distributions and falling back to a value sampling step when conjugacy fails. It identifies closed families (e.g., linear-Gaussian, finite discrete) that guarantee exact inference for annotated variables. Evaluation reports an average 1.6× slowdown versus prior delayed-sampling systems on existing benchmarks but shows 3×–87× speedups on new benchmarks that exploit closed-family exactness.
AI-written summary · about
chapters (11)
- 0:00Semi-symbolic inference intro
- 1:23Hybrid representation in filters
- 3:30Expression trees and swaps
- 4:56Streaming PP example
- 6:43Runtime for semi-symbolic inference
- 11:16Approximate sampling with values
- 12:39Closed families and guarantees
- 14:02Empirical performance gains
- 15:50Related work overview
- 16:34Summary and contributions
- 17:04Q&A
related talks
19mInference Plans for Hybrid Particle Filtering
36mProbabilistic Programs Which Make (Common) Sense
19mIncremental Computation for Efficient Programmable Inference in Probabilistic Programs
31mProbabilistic Programming: The What, Why and How
35mScalable structure learning and inference for domain-specific probabilistic prog
20mProbabilistic Programming with Stochastic Probabilities
21mGuaranteed inference for probabilistic programs: a parallelisable, small-step o...
24mPractical Probabilistic Programming with Figaro