Asking the Crowd the Right Question: Bias-Cancelling Weights for Federated Learning
The paper argues that the choice of client weights is the central failure point in federated learning, because bad weights create bias no optimizer can remove.
The authors propose CROWD, a weighting method that estimates useful client error information from disagreement among clients during training. They say the method recovers per-client second moments through a stable linear inversion and can measure when realized estimates hit an incoherence floor. In their experiments, CROWD reaches oracle-level excess risk on real site-split scans with miscalibrated detectors, while simple noise-variance weighting can perform worse than uniform weighting. ArXiv · AI/CL/LG's note
The authors propose CROWD, a weighting method that estimates useful client error information from disagreement among clients during training. They say the method recovers per-client second moments through a stable linear inversion and can measure when realized estimates hit an incoherence floor. In their experiments, CROWD reaches oracle-level excess risk on real site-split scans with miscalibrated detectors, while simple noise-variance weighting can perform worse than uniform weighting. ArXiv · AI/CL/LG's note
score 3