Tutorial: Incentives for Collaborative Learning and Data Sharing, Part I
summary
Presents a formal model where K agents choose contribution levels (Mi) and utilities equal local model quality minus linear data cost, proves Nash equilibria exist but can be highly inefficient — an example requires log(1/ε) more samples at equilibrium than optimal. Empirical Federated Averaging runs show 94% of agents can cut to 25% of their data and remain satisfied. Reduction of the planner’s optimization to set cover makes exact optimal allocation NP‑hard, so VCG requires exact optimal solution and is infeasible; LP approximations plus pay‑as‑contribute rules are the practical alternatives.
AI-written summary · about
chapters (8)
- 0:00Welcome
- 1:29Incentivizing data sharing
- 8:49Desirable properties
- 17:20Efficiency and rationality
- 20:25Monetary payments in PAC learning
- 32:10Central planner mechanism
- 42:32Payment mechanism limits
- 44:27Break
related talks
48mPersonalized Collaborative Learning with Affinity-Based Variance Reduction
48mIncentivizing Truthful Data Sharing in Collaborative Machine Learning and Data Marketplaces
44mPlatforms for Efficient and Incentive-Aware Collaboration
42m"Federated Reinforcement Learning: Statistical and Communication Trade-offs
30mLearning Multi-Agent Collaborations With Decomposition
45mAGT and Machine Learning
01h 28mMulti-Distribution Learning, for Robustness, Fairness, and Collaboration
01h 07mIndustry Presentations