If We Want AI to be Interpretable, We Need to Measure Interpretability
summary
Argues that interpretability must be measured, not promised, and introduces two concrete approaches: an intruder interpretability metric for topic models that reveals a surprising negative correlation between held-out likelihood and human coherence, and a bandit-based explanation optimizer for question answering that learns which explanations boost human accuracy. Reports human studies where novices gain most from evidence pages while experts gain from highlighted cues, shows tailored explanations beat showing everything and beat the AI alone, and finds a ~9-word wait optimizes simultaneous translation for QA.
AI-written summary · about
chapters (5)
- 0:00Intro
- 1:23AI Should be Interpretable
- 12:12We Should Measure Interpretability
- 26:02Proposal for Unsupervised Methods (Topic Models)
- 31:29Proposal for Supervised Methods (Question Answering/Translation)
related talks
28mThe Need for Explainable AI
32mOn the Importance of Explainability
31mAccuracy and Interpretability Through the Lens of Human-AI Teaming
28mA Practitioner’s Guide to Opening the AI Black Box for Humans
45mAI-404: Explanation Not Found
20mWhat can AI do for me?: evaluating machine learning interpretations in cooperative play
21mWhy you should care about AI interpretability
26mUsing Explainable AI to garner trust and overcome adoption inhibitors