Using Spiking Neural Networks for Control Applications at the Edge

Abstract

In recent years, the automation of vehicles has been an area of growing interest. Everything from cars to planes is receiving more advanced technology to help it function autonomously. However, increasing levels of automation come at a cost, significant increases in computational complexity, and power consumption. Research has shown that Spiking Neural Networks, the third generation of neural networks, offer increased power efficiency and inference time through the exploitation of biological principles from recent breakthroughs in neuroscience. We set out to demonstrate that spiking neural networks could be used for control applications in Size, Weight, and Power (SWaP) constrained environments. We derived the physical dynamics of blimps and implemented them in a Python-based aerial flight simulator. Using spiking-based deep Q networks and Q-based reinforcement learning, we are training networks to identify, navigate towards, and capture surrounding targets using a single camera. Our work aims to highlight the ability to spike neural networks to control SWaP-constrained agents while remaining performant in spite of stochastic environments.