PushCen-ADFL — bias-corrected async federated learning without central server
New arXiv paper tackles asynchronous decentralized federated learning (ADFL): no central coordinator, but peer-to-peer gossip creates communication overhead, skewed aggregation, and model drift across non-IID data.
PushCen-ADFL couples compression, aggregation, and local stabilization in a shared centroid space — closing the loop between what clients exchange and what the optimizer sees, even under delayed updates and asymmetric links.