Split Learning in Wireless Networks: A Communication and Computation Adaptive Scheme

Abstract

By deploying deep learning tasks between the mobile devices and the edge servers collaboratively, split learning provides a feasible method to fully integrate dispersed computation resources at the edge of wireless networks. However, due to the high dynamics of wireless networks, it is challenging to balance the cost and the computation efficiency. To satisfy extreme user experience requirements of intelligent-enabled applications, a communication and computation adaptive scheme is studied in this paper to achieve high efficiency with low costs. First, an adaptive split learning paradigm is designed to support flexible management of model splitting and computation resources, which can balance communication and computation in dynamic wireless circumstances. Second, a deep R-learning network based algorithm is proposed to make the instantaneous decision for the long-term average cost minimization, by accounting for the undis-counted average cost and the curse of dimensionality. Finally, the simulation results are provided to show the performance gains of our proposed algorithm.