Multi-objective Learning: An Algorithmic Toolbox for Optimal Predictions Anytime Anywhere!
summary
Presents a concrete algorithmic toolbox that turns K-fold hard multi-distribution learning into an adaptively sampled problem where sample complexity adds rather than multiplies: allow on-demand sampling and adaptive re-sampling to go from ≈K×n to roughly n + polylog(K)·n (realizable bounds sketched), use game reductions with no-regret vs best-response dynamics to control rounds vs samples, and construct learnable panpredictors (omnipredictor generalizations) that post-process into optimal predictors for many losses; experiments show resample+reweighting outperforms pure reweighting.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:27Origins of multi-objective learning
- 4:42Multi-objective framework
- 9:04Fine-grained multi-objective learning
- 13:24Algorithmic toolbox and research
- 16:41On-demand sampling
- 21:40Game-theoretic approach
- 25:16Sample complexity strategies
- 30:56Panpredictors and Bayes predictors
- 34:50Summary and future
- 35:12Q&A
related talks
01h 11mA Theory of Multi-objective Machine Learning
01h 28mMulti-Distribution Learning, for Robustness, Fairness, and Collaboration
48mOnline Adversarial Multicalibration And (Multi)Calibeating
34mPandora's Box with Correlations: Learning and Approximation
01h 21mMulticalibration: Towards Fair Decision Making
47mA Complex Picture of Multi-task Learning
32mDesigning Adaptive Experiments For Policy Learning And Inference
37mOmnipredictors