Biases are bugs: algorithm fairness and machine learning ethics
summary
Claim: biases are bugs that must be found, fixed, and learned from. Concrete methods include stress cases that enumerate worst harms, automated association testing with Erica's package to flag protected‑class correlations, and feature distribution repair that remaps percentiles to reduce disparate impact shown to move metrics from <0.8 to ≥0.8. Worked examples also include rotating and debiasing word embeddings so analogies stop encoding stereotypes. Tradeoffs between fairness and performance are measured and sometimes cost only a percent or two.
AI-written summary · about
chapters (12)
- 0:00Algorithmic bias introduction
- 3:18Detecting algorithm biases
- 6:37Association tests for discrimination
- 12:01Removing biases via data engineering
- 17:00Fixing biased word embeddings
- 21:02Building fair algorithms
- 27:27Q&A: Dataset repair trade-offs
- 30:37Q&A: Fairness code of conduct
- 32:45Q&A: Intersectional bias detection
- 35:12Q&A: Bias and model selection
- 36:14Q&A: Bias correction features
- 37:27Closing remarks
related talks
31mFlagging *and* Fixing Bias in ML
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....
51mDigital Discrimination: Cognitive Bias in Machine Learning
25mIt's not a bug, it's a bias
56mAn AI with an Agenda: How Our Cognitive Biases Leak Into Machine Learning
47mFight AI Bias With… Bias
38mSocial Implications of Bias in Machine Learning
47mEthical Machine Learning