Can AI design a fairer future?
summary
Uses the COMPAS recidivism example and an insurance pricing use case to show a concrete adversarial technique: an encoder H with HGR penalty that decorrelates a latent Z from a sensitive variable so predictions contain no recoverable gender signal. A colored-digit experiment (10 digits with color confound) reports baseline test accuracy ~0.447 and shows the method outperforms a post hoc remover. Code and evaluation live on GitHub as furnace Compass. Warns that fairer models require access to sensitive labels to enforce fairness.
AI-written summary · about
chapters (9)
- 0:00AI design for fairer future
- 0:30COMPAS algorithm bias
- 2:15Machine learning bias reproduction
- 4:42Mitigating algorithmic bias
- 6:49Neutral AI model representation
- 8:12Testing fairness example
- 10:19Defining fairness limits
- 11:22Fairness compass decisions
- 12:04Ethics of sensitive data access
related talks
23mRemoving Unfair Bias in the Machine Learning Pipeline
55mEthical Machine Learning
21mFairness and bias in AI
54mEthical Machine Learning
34mInterpretability and Algorithmic Fairness
47mEthical Machine Learning
27mCrouching Harm, Hidden Bias—Addressing issues of fairness....
46mA look at the methods to detect and try to remove bias in ML models