Making Logic a First-Class Citizen in Generative ML for Networking
summary
NetNomos learns first-order logic rules from four real-world datasets, filters them, then enforces them at inference by running a GPT-2 model alongside an SMT solver to constrain outputs just-in-time. The system is 1.6–6.5× more scalable than DuoAI at rule learning, enforces zero rule violations across telemetry imputation, traffic forecasting, and synthetic trace generation, and matches or beats specialized systems like Zoom2Net and NetShare on Meta’s data-center traces while preserving statistical fidelity.
AI-written summary · about
chapters (11)
- 0:00NetNomos introduction
- 1:44Logic rule violations in AI
- 2:48Challenges in network logic rules
- 5:38NetNomos core approach
- 6:22Rule learning via constraints
- 8:12Learning expressive network rules
- 10:36Rule enforcement during inference
- 13:25Evaluating with real-world data
- 14:06Experimental results
- 16:59Summary and release
- 17:20Q&A
related talks
29mSwitchLog: A Logic Programming Language for Network Switches
15mCurrent Practices for Building LLM-Powered Reasoning Tools Are Ad Hoc—and We Can Do Better
43mKeynote: Applied NLP in the age of Generative AI
54mDeterministic GenAI Outputs With Provenance
26mLanguage Model Agents Enable Semi-Formal Programming
22mGenerative AI - A Pragmatic Business View
26mGrappling with the Bitter Lesson
16mMulti-modal Program Inference: a Marriage of Pre-trained Language Models and Component-based Synthes