A Complexity-Theoretic Perspective on Fairness
summary
Defines multi-calibration as calibration not just overall but on subpopulations specified by a computational class, and shows how to reduce learning such predictors to standard ML tasks; introduces outcome indistinguishability by turning a predictor into a generative model (resample Y by Bernoulli(p̂(x))) and requiring no efficient distinguisher class A can tell model from Nature, i.e. A-outcome indistinguishable; proves tight connections between OI and multi-calibration and reports applications to ranking, robust estimation, outlier detection, and agnostic learnability.
AI-written summary · about
chapters (6)
- 0:00Intro
- 1:28Predictive algorithms everywhere you look!
- 2:45A reactive approach to fairness Addressing the blases the public has seen
- 4:14A foundational approach to fairness Towards responsible machine learning
- 6:57Outcome Indistinguishability
- 9:58Connections Beyond fairness
related talks
47mA Multi-Group Perspective on Fairness
50mA Multi-Group Approach To Algorithmic Fairness
01h 21mMulticalibration: Towards Fair Decision Making
45mMulticalibration, Universal Adaptability and Causality
37mOmnipredictors
53mGroups and Individuals in Fairness and Probability
24mLoss Minimization through the lens of Outcome Indistinguishability
43mAfternoon Session: Individuals, Groups, Pattern Recognition and Rich Data