Federated Stochastic Configuration Networks with Universal Approximation Property

Stochastic Configuration Networks (SCNs), a class of randomized learner models, are incrementally constructed under a supervisory mechanism, ensuring the universal approximation property. As powerful tools for data modeling, SCNs have received considerable attention due to their guaranteed reliable learning performance, easy implementation, and robustness with respect to learning parameters. In this work, we aim to touch the base of Federated Learning (FL) with neural networks, and propose a novel FL framework with SCN models, termed Federated Stochastic Configuration Networks (FSCNs). Our key contribution lies in the first time establishing some fundamental results on the universal approximation property, which can be regarded as a milestone in the working field. The proposed method incorporates a collaborative Particle Swarm Optimization (PSO) mechanism, enabling clients to collectively optimize hidden nodes without sharing local data. Unlike the traditional Federated Averaging (FedAvg) methods, which rely on weighted average aggregation, our approach leverages the global search capability of PSO to configure new hidden nodes, thereby enhancing the learning performance and model stability. Experiments on both benchmark and real-world datasets show that the proposed Federated Stochastic Configuration Networks (FSCNs) outperform the existing solutions in terms of learning capability, convergence speed and model compactness.