Nowadays, cyber-physical systems are omnipresent. Internet of Things (IoT) devices give physical sensing to computers. However, they also pose serious architectural challenges due to a number of reasons: they usually have different (and limited) computation power; they are manufactured by different vendors; and the battery consumption is a heavy limit. To overcome these problems, three main architectural solutions emerged: cloudlets, fog computing, and edge computing. All those paradigms enable the efficient handling of data from IoT sensors to the cloud and in particular so-called edge intelligence. Edge artificial intelligence (AI) is a term used in the past few years that refers to the confluence of machine learning and edge computing. Applications such as smart cars, smart cities, and the smart grid require heavy and fast data processing.

The paper presents a survey on edge AI. It starts by introducing the three main architectures and covering the main application scenarios (for example, IoT, security, autonomous driving) and their implementation. It is quite exhaustive in describing the projects and initiatives, both from academia and from industry (for example, with Azure or Amazon Web Services edge solutions). The authors then perform a deep dive on edge intelligence, describing machine learning algorithms at the edge, deep learning at the network edge, and their application scenarios. It also covers hardware and software architectures supporting edge intelligence.

This paper was written for a diverse target audience. First, it introduces edge and fog computing, specifically for readers who are not yet familiar with edge computing. These readers will benefit from discovering possible edge implementations. Second, researchers in edge AI will benefit from the pointers to various research projects underway in the field.