Probabilistic scripts for automating common-sense tasks

on YouTube ↗36m

" summary

Claims a declarative DSL called Pclean encodes domain knowledge as probabilistic generators, compiles to Gen, and runs programmable inference combining sequential Monte Carlo with expectation–maximization to both infer clean values and learn parameters from messy tables. It embeds in Julia, can call GPT‑2 for text priors, and reports matching state‑of‑the‑art accuracy on data‑cleaning benchmarks. The implementation is open‑source and the pipeline demonstrably resolves common typos, unit errors and missing values without hand‑coded heuristics.

AI-written summary · about

" chapters (12)

1. 0:00Introduction
2. 1:16Cleaning messy data
3. 4:18Writing the script
4. 7:42Declarative programming
5. 16:45Data cleaning
6. 18:25Encoding knowledge
7. 18:39Encoding domain knowledge
8. 20:45Pclean
9. 28:01GPT
10. 31:50probabilistic inference
11. 32:11PE architecture
12. 34:28Preliminary results

" related talks

• 35mScalable structure learning and inference for domain-specific probabilistic progFeras A. Saad · LAFI 2022
• 43mThe Power of Probabilistic ProgrammingBen Vigoda · O'Reilly Artificial Intelligence Conference 2017 - New York, New York
• 36mProbabilistic Programs Which Make (Common) SenseZenna Tavares · Strange Loop 2015
• 22mStochastic Lazy Knowledge Compilation for Inference in Discrete Probabilistic ProgramsVikash K. Mansinghka, Armando Solar-Lezama, Joshua B. Tenenbaum, Alexander K. Lew, Maddy Bowers · PLDI 2025
• 38mProbabilistic programming and meta-programming in ClojureVikash K. Mansinghka · Clojure/conj 2018
• 32mStructured Deep Learning with Probabilistic Neural ProgJayant Krishnamurthy · Scale By The Bay 2018
• 40mProbabilistic Programming for Augmented IntelligenceVikash K. Mansinghka · Uncertainty in Computation 2016
• 01h 03mAn Overview of Probabilistic ProgrammingVikash K. Mansinghka · Strange Loop 2015