Attributes: Selective Learning and Influence
summary
Characterizes which K attributes to sample when attribute realizations follow a Gaussian process and two players disagree on attribute weights, using Brownian, Ornstein–Uhlenbeck and squared‑exponential kernels; shows optimal sampling is governed by two sufficient statistics (principal’s posterior variance and the covariance between agent value and principal decision) and by a nearest-attribute property (Markov covariance). Under that property samples are mostly relevant (at most two idiosyncratic picks, rest common) and optimization breaks into k‑local problems; non‑Markov kernels can make sampling irrelevant attributes optimal when correlation is extreme.
AI-written summary · about
chapters (8)
- 0:00Introduction and context
- 2:00Model framework
- 8:00Gaussian process foundations
- 18:01Extrapolation and property
- 26:04Single-player benchmark
- 33:53Strategic sampling
- 41:11Conclusion and insights
- 49:07Closing remarks
related talks
01h 11mA Theory of Multi-objective Machine Learning
01h 02mTheory and Big Data
44mActive Regression via Linear-Sample Sparsification
35mLearning General Gaussian Mixtures With Efficient Score Matching
01h 22mGeneralization IV
38mMulti-objective Learning: An Algorithmic Toolbox for Optimal Predictions Anytime Anywhere!
53mDesigning Robust Learners
01h 28mMulti-Distribution Learning, for Robustness, Fairness, and Collaboration