Learning and Strongly Truthful Multi-task Setting Peer Prediction: A Variational Approach
summary
Defines a variational estimator of phi‑mutual information and uses it to build a multi‑task peer‑prediction payment: learn an ideal scoring rule K on withheld tasks, then pay K(reports on a common task) minus the convex conjugate term evaluated on reports from distinct tasks. Proves no unbiased phi‑MI estimator exists for finite tasks, shows the mechanism gives approximate strong truthfulness when K* is learned under truthful data, and handles continuous or arbitrary measurable signal spaces; it subsumes correlated‑agreement and Dasgupta–Ghosh as special cases.
AI-written summary · about
chapters (5)
- 0:00Start
- 1:31Peer Predictions
- 8:33Outline
- 9:07What Is the Mutual Formation Framework
- 9:29Prior Singular Test
related talks
47mA Complex Picture of Multi-task Learning
23mFalse Consensus, Information Theory, and Prediction Markets
48mIncentivizing Truthful Data Sharing in Collaborative Machine Learning and Data Marketplaces
52mMechanisms for Honest Data Sharing under Competition
51mRandom Projections for Probabilistic Inference
01h 11mA Theory of Multi-objective Machine Learning
01h 28mMulti-Distribution Learning, for Robustness, Fairness, and Collaboration
33mHeterogeneity and Privacy in Modern Learning