Abstract
In existing federated learning (FL), the base station (BS) coordinates devices to collaboratively train a shared model by avoiding the transmission of raw data. To achieve communication-efficient model uploading, over-the-air computation (AirComp) is often employed to aggregate model parameters. However, in conventional AirComp assisted FL, the BS's abundant computation resources are underutilized due to its non-involvement in model training. Meanwhile, transmission failures resulting from fluctuating wireless channels impair the quality of model aggregation. In this paper, we propose a retransmission-based semi-federated learning (SemiFL) framework, wherein devices upload model parameters and public privacy-free data for enabling a hybrid implementation of FL and centralized learning (CL). In our new framework, the BS leverages its abundant computation resources to aid CL model training, which mitigates the resource wastage while alleviating local computational burden of devices. Moreover, the proposed new retransmission mechanism effectively overcomes detrimental transmission failures resulting from the fluctuating quasi-static channel, aiming to guarantee improved learning performance of SemiFL. Successful transmission probabilities of both retransmission-based AirComp and retransmission-based digital communication are provided in closed forms. To attain deep insights, we derive an optimality gap to capture the convergence behavior of retransmission-based SemiFL. Then, we formulate a non-convex long-term problem to minimize a weighted sum of overall latency and energy consumption by jointly optimizing communication, computation, and learning parameters. Extensive experimental results show that our retransmission-based SemiFL obtains 21.9%, 30.5%, and 44.1% accuracy gains on three datasets, while efficaciously reducing latency and energy consumption compared to benchmarks. Meanwhile, our scheme enhances learning performance on the fluctuating quasi-static channel compared to state-of-the-art schemes.