Inclusive Storytelling With Data
summary
Shows concrete failures and fixes: a Google Translate demo flips “she is a doctor” to “he” because training corpora reflect historical gendered roles; COMPAS labels produce 45% vs 23% false high-risk rates for Black versus white defendants; self-driving vision misses dark-skinned pedestrians because training sets underrepresent them; Boston’s 2011 pothole app only maps rich neighborhoods until data collection moves to buses and garbage trucks. Recommends LIME for local explanations, Facets and What-If for dataset and model inspection, and making explainability a selection criterion.
AI-written summary · about
chapters (18)
- 0:00Introduction
- 0:48Douglas Adams
- 1:16Babel Fish
- 1:58Demo
- 3:36Cancer
- 4:40Indigo
- 8:29Artificial Intelligence
- 12:01Machine Learning
- 13:33City of Boston
- 19:29Compass System
- 23:10SelfDriving Cars
- 24:54Model Selection Criteria
- 25:37Visualization
- 25:52Lime
- 26:18Agnostic
- 27:26Diversity
- 28:24Remove Data
- 29:21What If
related talks
38mSocial Implications of Bias in Machine Learning
38mSocial Implications of Bias in Machine Learning
47mEthical Machine Learning
37mThe AI 'Black Box'
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....
41mEthics Panel
28mBreaking the Chains of Oppressive Software
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World