Learning to communicate (LeanCom): Deep learning based solutions for the physical layer of communications
The talk presents an overview and technical highlights of project LeanCom “Learning to Communicate: Deep Learning based solutions for the Physical Layer of Communications” on AI-inspired physical layer wireless communications. The basis of the research is the advancement of signal processing for physical layer wireless communication design, to address complex scenarios where classical signal processing approaches underperform. This is done by departing from traditional approaches based on theoretically inspired block-by-block transceiver design, towards practically oriented data-driven and model-driven approaches. Such a transmission entails an number of fundamental changes in the design of wireless links: a) Design of neural networks (NNs) tailored to communications applications, b) new transceiver architectures through the use of these communications-specific NNs and layered training mechanisms tailored to the connectivity scenarios, c) hardware-informed and hardware-efficient NNs suitable for practical deployment in the dynamic communications scenarios.