Statically and Dynamically Delayed Sampling for Typed Probabilistic Programming Languages
summary
Implements statically delayed sampling as a compile-time transform for Miking CorePPL that builds programmatic Bayesian networks, applies local marginalization and conditioning, and regenerates optimized code to remove runtime relation tracking. Claims: same marginalized log-likelihood as dynamically delayed sampling on LDA and Bayesian linear regression, while reducing execution time (static is consistently faster); limitation: static fails on models with recursion (epidemiology / vector-borne example), so dynamic remains necessary for those cases.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:50Bayesian inference basics
- 3:17Exact vs approximate inference
- 5:49Runtime overhead
- 8:40Compile-time delayed sampling
- 11:53Bayesian networks
- 13:56Conjugate prior transformation
- 16:07Type system challenges
- 18:56Evaluation metrics
- 21:50Conclusion
- 22:38Q&A
related talks
14mMulti-Language Probabilistic Programming
20mProbabilistic Programming with Stochastic Probabilities
01h 03mA Personal Viewpoint on Probabilistic Programming
29mMotivating Probabilistic Programming
23mDynamic inference in probabilistic graphical models
31mProbabilistic Programming: The What, Why and How
14mSequential Monte Carlo Program Synthesis with Refinement Proposals
26mSemi-symbolic Inference for Efficient Streaming Probabilistic Programming