Tutorial: Incentives for Collaborative Learning and Data Sharing, Part Ib
summary
Shows that naive collaboration can be K times inefficient: a double-star network has an optimal effort of 2 but any stable outcome costs Theta(K). Uses Bayesian persuasion to design a binary persuasive signaling scheme that reduces expected total effort to Theta(sqrt(K)) and proves that sqrt(K) is tight. For data fabrication, builds a leave-one-out style mechanism that adds noise scaled to an agent’s deviation so truthful reporting and contributing the socially optimal m are incentive-compatible*.
AI-written summary · about
chapters (12)
- 0:00Collaborative learning incentives
- 1:09Linear model for collaborative learning
- 5:22Platform role in collaboration
- 9:55Efficiency vs stability
- 14:56Information design and Bayesian persuasion
- 27:56Beyond data fabrication
- 37:13Incentive-compatible truthful sharing
- 48:11Tutorial summary and challenges
- 50:07Q&A on literature and costs
- 55:26Real data market incentives
- 58:33Protocol timeline and beliefs
- 59:15Open problems and future work
related talks
44mPlatforms for Efficient and Incentive-Aware Collaboration
48mIncentivizing Truthful Data Sharing in Collaborative Machine Learning and Data Marketplaces
52mMechanisms for Honest Data Sharing under Competition
01h 28mMulti-Distribution Learning, for Robustness, Fairness, and Collaboration
14mInformation, Learning, And Incentive Design For Societal Networks
30mLearning Multi-Agent Collaborations With Decomposition
48mPersonalized Collaborative Learning with Affinity-Based Variance Reduction
45mAGT and Machine Learning