Hardware Implementation for Spiking Neural Networks on Edge Devices

Abstract

In the modern data-intensive Internet of Things (IoT) applications, intelligence is enabled by collecting data at the edge devices for processing. The general lack of data processing power in these devices (due mainly to resource constraints) means that the data collected needs to be transmitted to the cloud for processing. The key issues arising from the use of the data communication channel include large energy consumption, communication delays and bandwidth issues, and potential security risks. An attractive approach to avoid these issues is to imbue the edge devices with more powerful data processing capabilities. The main design challenges to improve the data processing capabilities of edge devices are associated with the tight energy and hardware resource constraints in these edge devices. Moreover, the neural networks that are widely deployed for IoT applications demand a large amount of memory and energy. In this chapter, we review the techniques that aim to improve the data processing capabilities of edge devices. Generally, the research community has proposed techniques aimed at improving the algorithms as well as the underlying hardware. We will first discuss the artificial neural network (ANN) and the spiking neural network (SNN) to motivate our focus on hardware design techniques for emulating SNN on the edge devices. Thereafter, this chapter will present a survey of the hardware architecture and optimization techniques proposed to emulate the state-of-the-art SNN on the edge devices.