Personalized Collaborative Learning with Affinity-Based Variance Reduction
summary
Proposes AffPCL, a personalized collaborative algorithm for multi-agent linear systems that uses bias correction and importance correction to handle objective and environment heterogeneity; proves sample complexity improves by a factor of max{n^{-1}, δ} with δ∈[0,1] (so when δ≈0 it recovers 1/(nT) federated linear speedup, and when δ=1 it reduces to independent learning 1/T); casts central-objective and density-ratio estimation as multi-agent linear problems and shows some agents near the virtual center can free ride and gain linear speedup—validated in simulations (Affinity-based variance reduction, density-ratio estimation).
AI-written summary · about
chapters (10)
- 0:00Personalized collaborative learning
- 1:01Talk overview
- 2:37Problem setup
- 7:57Goals and baselines
- 13:14Key results
- 18:09Heterogeneity and affinity
- 28:01High heterogeneity regime
- 28:47Algorithm design
- 44:38Future research
- 46:38Q&A
related talks
46mEliciting and Enforcing Subjective Individual Fairness
45mTutorial: Incentives for Collaborative Learning and Data Sharing, Part I
30mIndividual Statistical Fairness
35mOnline Learning with A Lot of Batch Data
47mA Multi-Group Perspective on Fairness
43mIs Interaction Necessary for Distributed Private Learning?
39mA PAC Approach to Application-Specific Algorithm Selection
30mLearning Multi-Agent Collaborations With Decomposition