The effects of example-based explanations in a machine learning interface
summary
A 2x2 randomized deployment to 1,150 QuickDraw users compares normative explanations (random samples from the true class) and comparative explanations (nearest‑neighbor embeddings from other classes) for a sketch recognizer. When the model fails, normative examples raise perceived system capability and user understanding; comparative examples often provoke surprise, expose limitations, and increase viewing time, producing no clear trust benefit. The paper reports these measured differences and recommends choosing example types based on whether the goal is reassurance or surfacing failure modes.
AI-written summary · about
chapters (9)
- 0:00Introduction to research
- 1:31Example-based explanations overview
- 2:37Study setup in sketch recognition
- 5:08In-the-wild user study design
- 6:58Results on system capability
- 8:26Comparative vs normative differences
- 9:31Key findings summary
- 10:13Future of human-machine collaboration
- 10:55Q&A session
related talks
48mInherent Trade-offs with the Local Explanations Paradigm
16mBuilding Trust through Explainable AI
45mExplainable Machine Learning Explained
17mChallenges Towards Production-Ready Explainable Machine Learning
39mKeynote 3: How did you know? Explaining Black Box Model Predictions in Python
37mExplaining model explainability
09mQuantitative Evaluation of Machine Learning Explanations: A Human-Grounded Benchmark
15mIntroduction to Explainable AI