Accuracy and Interpretability Through the Lens of Human-AI Teaming
summary
Argues that treating models as components of human-AI teams flips the accuracy versus interpretability trade-off: the most predictive model can harm system performance when humans and models have complementary strengths. Uses group decision literature, per-region loss analysis, deferral and confidence scores, clustering of input space, and the Rashomon set idea to show how choosing models by average accuracy misses complementarity and dynamic effects; proposes changing objectives, deferral rules, and model selection to optimize team-level outcomes rather than model-only metrics.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:45Rethinking decision support
- 5:40The standard view limitations
- 11:54Complementarity in teams
- 17:08Interpretability as a means
- 22:12Deep-level complementarity
- 26:57Q&A and closing
related talks
31mDecision theoretic foundations for human-AI collaboration
41mIf We Want AI to be Interpretable, We Need to Measure Interpretability
40mHuman-AI Collaboration Models for Better Decisions and Faster Workflows
13mHuman + Machine
46mCan you trust your AI?
31mA more human-centered approach to AI
37mInteractive Explainable AI
45mAgreement and Alignment for Human-AI Collaborative Decision Making