Scaling Exact Inference for Discrete Probabilistic Programs
summary
Introduces Dice as a discrete PPL that compiles programs to binary decision diagrams and performs inference by weighted model counting, reusing compiled function BDDs to amortize cost. Claims exact inference scales to hundreds of thousands of discrete variables (examples: Caesar-cipher inference to ~10,000 characters, Munin Bayesian network with >100,000 variables) and reports seconds-level runtimes; source, paper and demo at dicelang.cs.ucla.edu.
AI-written summary · about
chapters (10)
- 0:00Intro: What Is Probabilistic Programming?
- 2:47Motivation: Caesar Cipher
- 3:48Caesar Cipher Program . Observed ciphertext: TVSK
- 5:48Challenge of Inference: Rejection Sampling
- 7:30Challenge of Inference: Path Enumeration
- 8:17Factorized Inference
- 10:58Summary
- 12:26Bayesian Network Scaling • More experiments in the paper
- 14:37Conclusion & Future Work
- 15:02Questions?
related talks
14mMulti-Language Probabilistic Programming
11mProbabilistic Programming Meets Automata Theory: Exact Inference using Weighted Automata
19mIncremental Computation for Efficient Programmable Inference in Probabilistic Programs
11mExact Inference for Discrete Probabilistic Programs via Generating Functions
05mBit-Blasting Probabilistic Programs
35mScalable structure learning and inference for domain-specific probabilistic prog
24mCompiling with Generating Functions
14mExact Probabilistic Inference Using Generating Functions