Envisioning Physical Layer Flexibility Through the Power of Machine-Learning

This paper presents the vision of an adaptive radio frequency (RF) communication signal processing pipeline solely composed of machine learning domain operations, aiming to provide a fully hardware-accelerated alternative to dedicated RF chips. A hybrid architecture, comprising elements of classic signal processing and learnable algorithms trained in an end-to-end manner, is proposed, that is compatible with contemporary ML hardware accelerators. The RF -ML pipeline, including speed-up optimization modifications, are explained in detail, followed by a brief description summarizing the deployment workflow of the end-to-end system on a pair of AI-enabled space-grade FPGAs. Finally, a bit-error- rate performance study of the simulated system as well as a HW -deployed setup including software-defined radios (SDR) validates the concept, followed by a detailed throughput benchmark over multiple AI-accelerator hardware configurations. Finally, we raise questions regarding practical implementation, such as receiver synchronization, restrictions of the ML-accelerator feature space, and weight quantization, which are discussed at the end of this paper.