Communication Technologies for Edge Learning and Inference: A Novel Framework, Open Issues, and Perspectives

With the continuous advancement of smart devices and their demand for data, the complex computation that was previously exclusive to the cloud server is now moving toward the edge of the network. For numerous reasons (e.g., applications demanding low latencies and data privacy), data-based computation has been brought closer to the originating source, forging the edge computing paradigm. Together with machine learning, edge computing has become a powerful local decision-making tool, fostering the advent of edge learning. However, the latter has become delay-sensitive and resource-thirsty in terms of hardware and networking. New methods have been developed to solve or minimize these issues, as proposed in this study. We first investigated representative communication methods for edge learning and inference (ELI), focusing on data compression, latency, and resource management. Next, we proposed an ELI-based video data prioritization framework that only considers data with events and hence significantly reduces the transmission and storage resources when implemented in surveillance networks. Furthermore, we critically examined various communication aspects related to edge learning by analyzing their issues and highlighting their advantages and disadvantages. Finally, we discuss the challenges and present issues that remain.

Introduction

The rapid growth in global data traffic is directly related to the accelerated popularization of edge devices. These typically low-powered embedded devices, which are often used for data collection, have led to unprecedented opportunities and innovative forms to improve our quality of life, serving as a stimulating substrate for new scientific discoveries [1]. Indeed, combining Internet of Things (loT) devices and data with recent breakthroughs in machine learning (ML) has suggested that academia and industry pursue solutions in scenarios related to smart cities, intelligent transportation, e-health, and e-banking. Particularly, ML thrives in the domains of these applications [2]. Typically, training procedures underlying ML models are computationally intensive; thus, only powerful cloud servers can support them effectively [3].