Panel on Societal Dimensions of Explanation
summary
Panel argues that interpretability must be tied to concrete aims and context, not lofty redefinitions: some defend a framework with contextual free parameters, others call a single general‑purpose solution a fool’s errand. Empirical work shows surprising results — people sometimes trust labeled black boxes more than simple decision trees — and common tools like LIME SHAP Grad‑CAM fail many sanity checks. Discussion stresses co‑design, measurement models and theory before wide deployment, and warns against rushing immature methods into production.
AI-written summary · about
chapters (14)
- 0:00Panel introduction
- 2:16Philosophy vs ML interpretability
- 9:37Interpretation factors
- 15:51Generic parameters for interpretability
- 18:04Empirical evidence in interpretability
- 27:50Model vs world explanation
- 35:22Dialogue-based interpretability
- 40:04Evaluating interpretability techniques
- 44:36Stakeholder trade-offs
- 48:13Hypothesis search in interpretability
- 54:41Sociological pressures
- 56:09Philosophical ML explanations
- 59:40Human-centered tools
- 1:06:03Closing remarks
related talks
26mStop Explaining Black Box Models, Use Interpretable Models Instead
45mPanel: The shape of AI governance
34mInterpreting Machine Learning Models: Why and How!
28mManaging AI Risk: Beyond Explainability
31mPanel Discussion
01h 29mNavigating ML Explainability: From Theory to Implementation
01h 36mOpen the Black Box: an Introduction to Model Interpretability with LIME and SHAP
16mBuilding Trust through Explainable AI