Scalable structure learning and inference for domain-specific probabilistic prog
summary
Claims an end-to-end pipeline that automatically learns domain-specific probabilistic programs for time series, tabular and relational data and compiles them to fast exact inference. Uses PCFG-style domain-specific languages, reversible-jump MCMC with involutive proposals, SMC/HMC and reverse-mode AD for parameters, implemented in Gen (Julia) and compiled to SPPL probabilistic circuits. Reports over 100x speedups versus prior PPLs and fairness queries answered in milliseconds (10x–3000x speedups); demos include airline change-point adaptation, zebrafish trajectories and 18M-cell relational datasets.
AI-written summary · about
chapters (11)
- 0:00Start
- 3:03Goal of Probabilistic Inference
- 4:13Problem-Solving Programming
- 9:31The Learning Problem
- 15:57Sequential Monte Carlo
- 20:13Domain-Specific Language for Relational Data
- 20:29Infinite Relational Model
- 22:15Sum Product Probabilistic Language
- 29:26Research Roadmap
- 33:31What Do Programs Look like
- 34:42End of the Session
related talks
14mMulti-Language Probabilistic Programming
19mIncremental Computation for Efficient Programmable Inference in Probabilistic Programs
22mStochastic Lazy Knowledge Compilation for Inference in Discrete Probabilistic Programs
36mProbabilistic scripts for automating common-sense tasks
41mScalable Collective Inference from Richly Structured Data using Probabilistic Soft Logic (PSL)
47mApproximate Lifted Inference with Probabilistic Databases
01h 03mAn Overview of Probabilistic Programming
40mProbabilistic Programming for Augmented Intelligence