Control-Inspired Federated Learning: A Projection-Based Approach

In this work, we study federated learning, a distributed learning framework where data remains decentralized across multiple clients, and a shared model is trained collaboratively via a central server. We take a control-theoretic approach, focusing on a specific class of residual neural networks by modeling them as dynamical systems. Building on this perspective, we propose FedProject, an algorithm designed to mitigate statistical heterogeneity caused by non-identically distributed data. Our method introduces a projection operator on the proximal term of FedProx and employs a two-step update to balance local learning and client drift from the global model. We evaluate FedProject against FedProx and FedAvg on both IID and heterogeneous datasets, demonstrating improved convergence and robustness to hyperparameter selection for this class of neural networks.