Scallop: A Language for Neurosymbolic Programming
summary
A Datalog-based language and compiler called Scallop combines a neural module for perception with a declarative logic layer and a differentiable provenance engine to learn intermediate structure from only end-to-end labels. It uses a relational data model, tracks uncertainty via provenance semirings, integrates with PyTorch and JAX, and reports a concrete result: a Static Pac-Man agent reaches 99.4% accuracy in 50 episodes versus an end-to-end CNN needing ~50,000 episodes.
AI-written summary · about
chapters (10)
- 0:00Neurosymbolic programming intro
- 3:01Defining neurosymbolic programming
- 7:21Scallop design choices
- 10:11Pac-Man demo results
- 17:07Scallop compiler architecture
- 23:11Provenance semirings
- 25:27Benchmarks and applications
- 30:49Foundation model integration
- 34:02Multimodal reasoning demo
- 35:08Q&A
related talks
52mI See What You Mean
17mBring Your Own Data Structures to Datalog
30mGraphs as the Backbone of Scalable Agentic Apps
17mInteractive Debugging of Datalog Programs
15marchive – Manohar Sai Jasti
48mPragmatic Code Generation for Efficient Execution
46mCoordination-Free Computations
52mData Over Code: A Short Hyperbole