Causality, robustness, and natural language understanding in ML
summary
Argues that recent gains come from huge web text and models trained at scale — compute up 3–4 orders of magnitude — but current systems still mostly pattern‑match under IID assumptions and hallucinate new facts. Proposes moving toward causal and modular representations from raw sensors, outlines concrete probes: translate A→B→A backtranslation, self‑supervised video forward/back consistency, and a controlled donation bot evaluation that measures real behavioral impact. Emphasizes causal representation learning and time‑series forward/back classifiers as open, practical directions.
AI-written summary · about
chapters (15)
- 0:00Introduction and welcome
- 1:32Advances in language models
- 5:35Challenges in conversational systems
- 8:09Limitations of current language models
- 11:10Imitation vs understanding in AI
- 13:49From structured to unstructured data
- 16:20Beyond IID assumptions in ML
- 17:46Causality in natural language
- 21:19Causal interactions in dialogue
- 23:06Disentanglement and causality
- 26:00Evaluating conversational AI
- 28:50Human diversity in evaluation
- 30:39Temporal graphs in dialogue
- 31:48Causality in time series and language
- 33:39Discourse and backward dependencies
related talks
06mCausal Machine Learning – What’s in it for Data Scientists?
41mMachine Learning can't do the thinking
01h 01mThe Mathematics of Causal Inference, with Reflections on Machine Learning and the Logic of Science
56mWhat can be passively learned about causality?
51mTowards a Pragmatic Understanding of Large Language Models
01h 11mThe Move Toward AGI: Why Large Language Models Surprised Almost Everyone... | Theoretically Speaking
29m🧠🧹 Causality - Mental Hygiene for Data Science
54mML Beyond Metrics: Challenges of Using ML Systems in the Real World