Scallop: A Language for Neurosymbolic Programming
summary
Scallop is a differentiable neurosymbolic language that compiles Datalog-style programs to a relational algebra machine and plugs into PyTorch, using a configurable provenance semirings framework to do efficient probabilistic and differentiable reasoning. Programs support Datalog with recursion, aggregation, negation and foreign functions and a library of 20+ “Providences.” Evaluated on eight tasks, Scallop trains a Pac-Man agent to 99.4% success in ~50 episodes (vs. DQN’s 50k+ for ~85%), beats GPT-3 by ~20% on Clutter, and keeps >90% accuracy with 25–50% data.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:27Neurosymbolic paradigms
- 3:35Pacman maze example
- 6:02Path planning in Scallop
- 8:29Learning paradigms in Scallop
- 10:55Performance benchmarks
- 12:58Conclusion
- 13:21Q&A: Learning Pacman
- 14:25Q&A: Interpretability
- 15:26Q&A: Debugging non-convergence
related talks
26mDSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines
14mProgramming Large Language Models with Algebraic Effect Handlers and the Selection Monad
24mRecursive Coding Agents
12mNeuroSync: Intent-Aware Code-Based Problem Solving via Direct LLM Understanding Modification
40mSurviving the Agentic Hype with Small Language Models
34mProgramming Languages for AI Programing Agents
28mTalking to Objects in Natural Language: Toward Semantic Tools for Exploratory(…)
15mCurrent Practices for Building LLM-Powered Reasoning Tools Are Ad Hoc—and We Can Do Better