Omnipredictors
summary
Shows that a multi-calibrated predictor (computed from data via a weak agnostic learner and multicalibration algorithms) serves as an omnipredictor for essentially all convex Lipschitz losses: a simple univariate post-processing function per loss matches or beats the best hypothesis in class C for that loss. Proof uses a partition-based multicalibration view and convexity; experiments and theory highlight a steep sample complexity bottleneck and computational cost for achieving multicalibration.
AI-written summary · about
chapters (10)
- 0:00Introduction to loss minimization
- 2:08Formalizing the paradigm
- 5:22The challenge of unknown losses
- 9:33Defining omni predictors
- 14:11Connection to multi-calibration
- 17:13The main theorem and benefits
- 20:31Interpreting multi-calibration
- 24:28Loss-free learning perspective
- 29:43Proof sketch and summary
- 34:05Future directions and conclusion
related talks
01h 21mMulticalibration: Towards Fair Decision Making
24mLoss Minimization through the lens of Outcome Indistinguishability
47mA Multi-Group Perspective on Fairness
45mMulticalibration, Universal Adaptability and Causality
12mA Complexity-Theoretic Perspective on Fairness
48mOnline Adversarial Multicalibration And (Multi)Calibeating
50mA Multi-Group Approach To Algorithmic Fairness
40mOmnipredicting Single-Index Models with Multi-Index Models