Inference Plans for Hybrid Particle Filtering
summary
Built Siren, a probabilistic programming language that accepts inference plan annotations to force hybrid particle filtering to keep some random variables symbolic and sample others, and an abstract-interpretation satisfiability analysis that proves whether annotations can always be implemented. Evaluation across three hybrid filtering algorithms and benchmarks shows average 1.76x speedup and up to 206x to reach a target accuracy, average 1.83x accuracy improvement and up to 595x, and the static analysis finds all satisfiable plans in 27 of 33 settings.
AI-written summary · about
chapters (11)
- 0:00Hybrid particle filtering intro
- 2:33Exact vs approximate inference
- 3:58Hybrid inference methods
- 5:23Choosing inference plans
- 6:29Inference plans interface
- 9:16Satisfiability analysis
- 11:22Siren implementation
- 11:44Evaluation of inference plans
- 13:08Satisfiability precision
- 14:09Conclusion and future
- 14:53Q&A
related talks
26mSemi-symbolic Inference for Efficient Streaming Probabilistic Programming
14mSequential Monte Carlo Program Synthesis with Refinement Proposals
24mPractical Probabilistic Programming with Figaro
47mApproximate Lifted Inference with Probabilistic Databases
31mProbabilistic Programming: The What, Why and How
32mProbabilistic and Differentiable Programming in Scientific Simulators
20mProbabilistic Programming with Stochastic Probabilities
21mProbabilistic Programming with Programmable Variational Inference