Reconfigurable Intelligent Surface for Green Edge Inference in Machine Learning

To provide energy-efficient machine learning services for high-stake applications, e.g., drones and autonomous cars, in this paper, we propose a reconfigurable intelligent surface (RIS)-empowered edge inference architecture by cooperatively executing tasks at multiple computing-enabled base stations. Specifically, an RIS with many reflecting elements is deployed in the network architecture to assist the controllable signal propagations via configuring the phase shifts of these elements, thereby enhancing the signal quality at receivers. To minimize the power consumption for RIS-empowered edge inference process, we shall propose a joint group sparse beamforming and phase shifts design approach for inference tasks allocation and signal propagations control, respectively. An alternating minimization method is further developed to decouple the optimization variables and split this intractable problem into two subproblems. The mixed l(1,2)-norm and difference-of-convex-functions (DC) techniques are presented respectively for group sparsity inducing and phase shifts design. Simulation results demonstrate the admirable performance gains of the proposed algorithms and the effectiveness of the deployment of RIS.