Machine Learning Fairness: Lessons Learned
summary
Concrete lessons on ML fairness focus on data, measurement, and design: open datasets need geographic and contextual coverage (Open Images Extended is 80% from India), models need slice-level metrics and synthetic or crowdsourced augmentation (Quick Draw shoes, Google Clips families, Jigsaw Perspective moved from Toxicity V1 to V6 using synthetic examples), and interfaces must surface ambiguity (Translate returns both amigo/amiga). Tooling commitments include Open Images Extended, Fairness Indicators, What-If, Model Cards, and a live Kaggle dataset for real comments.
AI-written summary · about
chapters (15)
- 0:00Introduction
- 1:42Human Design
- 2:35Why We Care
- 5:38Welcome
- 6:26Data
- 11:10Measurement
- 15:12Improvement
- 26:55Transparency Frameworks
- 27:20Transparency Framework 1
- 30:48Real Comments
- 31:24WhatIf Tool
- 32:34Fairness Indicators
- 34:26Case Studies
- 35:02Commitment
- 35:22Conclusion
related talks
42mFairness in Machine Learning: From Theory to Practice
37mKeynote: Ethical Machine Learning: Creating Fair Models in an Unjust World
10mWhy diversity matters in Machine Learning communities
40mResponsible AI Practices: Fairness in ML
35mBuilding Inclusive Machine Learning Models
41mMachine Learning Model Fairness and Interpretability with AWS
38mSocial Implications of Bias in Machine Learning
47mEthical Machine Learning