Panel Discussion
summary
Panelists argue that the core issues are practical: using mutual information or max-entropy to justify coarse-graining produces misleading identification, and treating a macro intervention as an average over microstates is wrong — macro interventions must map to specific micro states. Examples include robots with different sensors, contextual bandits lacking historical policies, and cholesterol effects hiding HDL/LDL confounding. Offline datasets often miss crucial context or policy information — offline data often lacks full context or policy. Identifiability methods like VAEs break when latent causal relations change — identifiable VAEs fail when causal relations change.
AI-written summary · about
chapters (11)
- 0:00Hybrid panel introduction
- 1:03Hypercontractivity and mutual information
- 3:19Offline vs online data asymmetry
- 6:12Parametric vs non-parametric balance
- 9:26Causal learning assumptions
- 14:07Identifiable VAEs limitations
- 16:37Critique of coarsening assumptions
- 20:17Macro variables purpose
- 23:36Causal relations from macro data
- 26:03Soft interventions in causal models
- 29:40Open questions and future work
related talks
01h 06mPanel on Societal Dimensions of Explanation
49mPanel Discussion
41mKeynote: Personalization, Bandits, and Causal inference
30mTalk by Suhas Diggavi (UCLA)
35mOnline Learning with A Lot of Batch Data
44mPanel: Building responsible, human-centric AI models
45mPanel: The shape of AI governance
28mManaging AI Risk: Beyond Explainability