What can be passively learned about causality?
Talkat Understanding Higher-Level Intelligence from AI, Psychology, and Neuroscience Perspectives · 2024-06-24..06-28·
summary
Argues that models trained by behavioral cloning on expert interventional data can learn actionable causal strategies: a simulated agent with a five‑variable latent DAG, trained on expert traces, learns to intervene and generalize to held‑out graphs at test time. Explanations and reasoning traces make features salient and let passive learners generalize out-of-distribution—even from perfectly confounded training—and a pure LM (Chinchilla 70B) prompted with few‑shot examples plus explanations and chain‑of‑thought learns intervention strategies. passive learning interventional data
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:03Causality and RL agents
- 4:06Passive learning limitations
- 8:14Routes to causal understanding
- 10:22Causal DAG experiments
- 15:41Role of explanations
- 21:07Shaping generalization
- 22:20Language models and causality
- 24:53Abstraction and conclusion
- 29:41Q&A and discussion
related talks
33mLearning Abstract Causal Structure From Causal Interventions In Children And Artificial Agents
28mLearning Causal Representations From Unknown Interventions
29m🧠🧹 Causality - Mental Hygiene for Data Science
34mCausality, robustness, and natural language understanding in ML
06mCausal Machine Learning – What’s in it for Data Scientists?
40mThe causality trap: why your LLM sounds smart but thinks shallow
41mOn the Path to Causal Inference
01h 01mIntroduction to Experimental Design