Abstract
In cellular-based federated learning (FL), the base station (BS) is only used to aggregate parameters, which incurs a waste of computing resources at the BS. In this paper, a novel semi-federated learning (SemiFL) framework is proposed to break this bottleneck, where local devices simultaneously send their gradient updates and training samples to the BS for global model computation. To capture the performance of SemiFL over wireless networks, a closed-form convergence upper bound of SemiFL is derived. Then, a non-convex problem is formulated to improve the convergence behavior of SemiFL, subject to the transmit power, communication latency, and computation distortion. To solve this intractable problem, a two-stage algorithm is proposed by controlling the transmit power and receive beamformers. Numerical experiments validate that the proposed SemiFL framework can effectively improve accuracy and accelerate convergence as compared to conventional FL.