Artificial intelligence in edge devices

Abstract

In the current era, advancements in deep learning have seen the services and uses of artificial intelligence (AI) flourish. From individual support and commendation structures to video/audio monitoring, there is something for every person. With the development of mobile computing and the Internet of Things (IoT), billions of mobile and IoT devices are now connected to the Internet, generating tons of data at the edge of the network. As a result of this inclination, there is a demanding need to push AI beyond its limits to the networks edge in order to fully understand the capability of big data. Edge computing, a developing model that encourages computing everyday jobs as well as network services, which are basic to the edge of the network, has long been touted as a hopeful explanation for meeting this demand. The resulting new interdisciplinary field known as edge AI or edge intelligence (EI) has been attracting an incredible amount of attention. Although EI experimentation is still in its early stages, both the computer system and AI societies would benefit from a committed forum for exchanging recent EI accomplishments. At present, we are conducting a detailed survey of recent EI research events. We go over the context and enthusiasm along with inspiration for AI running first at the edge of the network.

Introduction

We live in the age of AI that has never before seen such rapid growth. Deep learning has been propelled forward by recent advances in algorithms, processing power, and large amounts of datasets [1]. Computer vision, speech recognition, and language processing, as well as chess (for example, AlphaGo) and robotics, have all seen major advancements owing to AI's remarkable reach [2]. A plethora of intelligent applications, such as intelligent personal assistants, personalized purchasing recommendations, video monitoring, and smart home appliances, have sprung up as a result of these advancements. These intelligent applications are well known for greatly enhancing people's lifestyles, increasing human work rate, and improving social productivity. Big data have lately seen a drastic change in the knowledge source from giant-scale cloud data centers to more ubiquitous end devices, such as mobile devices and Internet of Things (IoT) devices, as a significant driver that boosts AI development. Big data, such as online shopping records, social media-related items, and business information systems, have traditionally been created and mainly stored in large data centers. Cisco predicts that by 2021, people, machines, and things at the network edge will have created nearly 850 ZB [3]. Global data center traffic, on the other hand, will only reach 20.6 ZB by 2021. However, due to concerns about efficiency, cost, and privacy, driving the AI frontier to the sting ecology that lives at the bottom of the web is not trivial [4]. When transmitting a large amount of data across a wide area network (WAN), the pecuniary cost as well as the transmission time may be excessively high, making privacy leaks a major concern [5]. Another option is on-device analytics, which uses AI programs to process IoT data directly on the device, but it could suffer from poor performance and energy efficiency [6]. Recently, moving cloud services from the networks core to its edge, i.e., closer to IoT devices and data sources, has been proposed. A helper node can be any adjacent terminal devices capable of communicating with each other via device-to-device (D2D) communication [7], servers connected to access points (e.g., WLANs, routers, and base stations), network gateways, or even microdata gateways that can be accessed by neighboring devices. While edge nodes can vary in size from a credit card-sized computer to a microdata center with multiple server racks, the most important feature highlighted by edge computing is physical closeness to the information-generating sources. In essence, the physical proximity of information-processing and information-generating sources promises many advantages over traditional cloud-based computing paradigms, including low latency, energy efficiency, data protection, low bandwidth consumption, local awareness, and context [6], [8]. In fact, the combination of edge computing and AI has generated a new field of study known as edge intelligence (EI) or edge AI [9], [10]. Rather than relying solely on the cloud, EI leverages the best of widely available edge resources to provide AI insights. EI has gained a lot of interest from both industry and academia. For example, EI has been included in the well-known Gartner's advertising cycle as a new technology that could reach productivity levels in the next 5–10 years [11]. Pilot projects have been proposed by major corporations such as Google, Microsoft, Intel, and IBM to illustrate the benefits of edge computing in paving the way for AI. This initiative has enhanced various AI applications, from real-time video analytics to machine learning [12], reasoning support [13] to correct farming, smart homes [14], and industrial Internet of Things (IIoT) [15]. Notably, studies on and practice of this new interdisciplinary field of EI are still in their infancy stages. In both industry and academia, there is a widespread lack of a platform dedicated to reviewing, debating, and sharing the current advancements in EI. To close this gap, we perform a comprehensive and detailed assessment of the current EI research efforts in this study. First, we will go through the history of artificial intelligence. Next, the rationale, definition, and grading of EI will be discussed. Following that, we will categorize and discuss the evolving computer architectures and supporting technologies for EI model formation and inference. Finally, we will discuss about some of the open research problems and possibilities for EI. The following is a breakdown of this chapters structure.

• Section 2 provides an outline of AI's core ideas, focusing on deep learning—chosen AI's field.

• The purpose, definition, and rating of EI are discussed in Section 3.

• The architectures, enabling methodologies, systems, and frameworks for training EI models are discussed in Section 4.

• Section 5 discusses EI model inference structures, enabling methodologies, systems, and frameworks.

• EI's future directions and difficulties are discussed in Section 6. We believe that by conducting this poll, we will be able to evoke more interest, spark constructive debates, and inspire new research ideas on EI.