A Theory of Multi-objective Machine Learning
summary
Introduces a concrete formalism: learn one model good for k distributions (per-distribution losses and calibration), cast as a min‑max game solved by no-regret vs best-response dynamics and online bandit/full‑info tools; shows interactive, adaptive sampling cuts sample cost from naive linear‑in‑k VC blowup to roughly additive log‑H + poly(k) regret terms (practical algorithms and benchmark results reported); flags incentive failures (need individual rationality and stability for collaborators) and open gaps like whether the log k factor is necessary.
AI-written summary · about
chapters (7)
- 0:00Multi-objective ML intro
- 10:50Multi-objective learning framework
- 24:48Resource needs in multi-objective learning
- 34:30Adaptive learning protocols
- 46:19Offline problems with online algorithms
- 56:05Generalization with infinite hypotheses
- 1:03:11Incentives in resource procurement
related talks
01h 28mMulti-Distribution Learning, for Robustness, Fairness, and Collaboration
38mMulti-objective Learning: An Algorithmic Toolbox for Optimal Predictions Anytime Anywhere!
48mOnline Adversarial Multicalibration And (Multi)Calibeating
53mAdversarial Machine Learning and Instrumental Variables for Flexible Causal Modeling
01h 10mMulti-Dimensional and Non-Linear Mechanism Design and Approximation II
45mAGT and Machine Learning
04mOn gradient-based methods for finding game-theoretic equilibria
01h 21mAdversarial Bandits: Theory and Algorithms