Explanatory Learning: Unleashing New Knowledge in Unknown Languages
summary
Claims a machine can learn to read and write explanations by training two transformer modules—a conjecture generator and an interpreter—under the Explanatory Learning framework and testing them in a synthetic Odeen universe. Dataset: ~1,400 training phenomena with 1,000 examples each and 1,000 held-out phenomena. The conjecture+interpreter system discovers correct rules for >700/1,000 new phenomena versus ~200 for end-to-end learners, outputs interpretable symbolic rules, and even benefits from a learned interpreter that tolerates grammatical variation.
AI-written summary · about
chapters (8)
- 0:00Workshop introduction
- 2:56Project goals and context
- 7:13Bridging humans and machines
- 17:10Scientific learning methods
- 26:24Explanatory learning framework
- 37:51Model architecture and results
- 48:03Interpreters and implications
- 1:03:43Closing thoughts and Q&A
related talks
28mExplaining Machine Learning Models with LLMs
20mMechanistic Interpretability
16mBuilding Trust through Explainable AI
46mLanguage Models as World Models?
59mBetween the Layers– Interpreting Large Language Models
17mChallenges Towards Production-Ready Explainable Machine Learning
45mReasoning in Natural Language
24mOpen and Understandable: Designing Explainable AI for Everyone