Deriving an Efficient Processing Chain for Radar with Spiking Neural Networks

Artificial Intelligence has transformed radar signal processing, achieving unparalleled results in object detection and angle of arrival estimation. Yet, the energy demands of deep neural networks, especially recurrent Long Short-Term Memory networks, pose significant challenges for batterypowered devices such as automobiles, drones, and wearables. Brain-inspired neuromorphic computing promises groundbreaking energy efficiencies through sparse, eventbased computation and communication using spiking neural networks.

This work aims to create accurate and energy-efficient processing chains for radar, that utilize the benefits of neuromorphic computing. It delves into the exploration of neural networks architectures that range from classical networks over hybrid to fully spiking neural networks. Using gesture recognition as a case study, those network architectures were trained, evaluated, and
subsequently mapped to SpiNNaker 2 to assess their performance and energy costs. The results demonstrate the enormous potential of spiking but also hybrid neural networks for energy-efficient radar processing chains.