Conceptual Challenges In Connecting Interpretability And Causality
summary
Argues that connecting interpretability to causality requires explicit contextual assumptions and that prediction alone is insufficient. Demonstrates under-specification by training 100 BERT-style models with identical data and architecture but different random seeds, finding large differences in gender-dependence of sentence-similarity scores. Frames modern causal methods as interpreting fitted models via estimands, overlap/positivity checks, and sensitivity analyses for unobserved confounders, urging pipelines that report multiple compatible hypotheses rather than a single narrative.
AI-written summary · about
chapters (8)
- 0:00Introduction to speaker
- 1:42Two cultures of modeling
- 4:45Bridging prediction and theory
- 7:36Causality as interpretation
- 11:02The identification challenge
- 16:17Rashomon effects and models
- 19:00Under-specification in ML
- 34:00Contextualizing interpretations
related talks
41mIf We Want AI to be Interpretable, We Need to Measure Interpretability
21mInterpretability For Everyone
37mInteractive Explainable AI
26mCausal inference with Causallib
23mMargaret Mitchell, Senior Research Scientist, Google at MLconf Seattle 2017
50mDehumanizing Agents: Why Explainability is Crucial in the LLM Era
27mExplainable Artificial Intelligence and Continuous Delivery: Towards the intelligent organization
28mExplaining Machine Learning Models with LLMs