FedSteer: Taming Extreme Gradient Staleness in Federated Learning with Corrective Projections and Caching

FedSteer is a novel method introduced to address aggregation variance in federated learning caused by inconsistent client participation. It constructs a gradient subspace from cached client gradients, allowing active clients to project their gradients onto this subspace and guiding inactive clients' outdated gradients towards the current global objective. Experimental results indicate that FedSteer improves performance stability and accuracy by over 7% compared to baseline methods, making it a valuable technique for practitioners dealing with extreme gradient staleness in federated learning systems.