Unifying Logic and Probability: The BLOG Language
summary
Presents BLOG, an open-universe semantics probabilistic language that lets models create and delete objects via number statements, so identity and count are inferred rather than assumed. Uses MCMC over partial model structures and a compiler to generate model-specific inference code; compiled BLOG runs orders of magnitude faster than generic engines (700–1,000× vs BUGS, ~100M samples/sec on a laptop). Demonstrations: citation clustering down to ~1% error, seismic monitoring cuts missed events from ~30–50% to ~11–12%, and multi-target tracking and video applications succeed.
AI-written summary · about
chapters (16)
- 0:00Introduction to AI logic
- 1:01First order logic in AI
- 5:31Probability and uncertainty
- 7:42Combining logic and probability
- 12:18Open vs closed universe
- 16:12The blog language
- 20:26Building possible worlds
- 25:20Blog model restrictions
- 27:28Citation matching problem
- 34:01Inference and performance
- 37:29Applications and tracking
- 39:50Model-based queries
- 41:55MCMC inference engine
- 46:10Compiler speed advantages
- 50:03Nuclear monitoring application
- 56:36Future work in NLU
related talks
01h 03mAn Overview of Probabilistic Programming
38mProbabilistic programming and meta-programming in Clojure
01h 07mProbabilistic Programming and Readable Models
14mMulti-Language Probabilistic Programming
24mExact Bayesian Inference for Loopy Probabilistic Programs Using Generating Functions
31mProbabilistic Programming: The What, Why and How
39mGauguin, Descartes, Bayes: A Diurnal Golem's Brain
24mHigher order programming with probabilistic effects: A model of stochastic memo...