Edge Computing with Artificial Intelligence: A Machine Learning Perspective

The concept of IoT was proposed in 1999 for supply chain management, but now IoT covers a much wider area [17]. With the integration of IoT into traditional industries, many new application areas have been spawned, such as smart home, smart grid, smart traffic, and intelligent manufacturing. The idea of IoT is that things connected to the Internet form a huge network, achieving the interconnection of these things at any time and place. With the continuous development of IoT, the number of various sensors, smartphones, healthcare applications and online social platforms is soaring, and the resulting global data will increase to 175 zeta bytes (ZB) by 2025 according to the prediction of International Data Corporation (IDC) [18]. This huge data volume has facilitated the world of big data [19].

In the era of big data, the most direct and simple method for handling those data is to transfer the data to the cloud for processing. The annual global cloud IP traffic of 2016 was 6.0 ZB, and it is expected to reach 19.5 ZB in 2021, reported by Cisco in 2018 [20] . However, the computing power of the cloud is increasing linearly [21], which is much slower than the current rate of data growth. With the rapid growth of data, cloud computing will no longer be fully trusted.

There are some IoT application scenarios that require extremely fast response speeds. For example, in the scenario of intelligent driving, sensor devices such as cameras are installed in autonomous vehicles. These sensor devices can continuously obtain data from the surrounding environment during the autonomous driving mode. In the cloud computing model, these data will be uploaded to the cloud for computing, and the results will be returned back to the vehicle’s control chip. Considering the complicated driving environment of a vehicle, this method is actually very time-consuming, and it may even cause the smart vehicle to fail to make the right decision in a timely manner, resulting in serious consequences [3].

In the fields of augmented reality (AR) and virtual reality (VR), mobile AR/VR applications need to continuously transmit high-resolution videos, so they have high requirements for data computing capabilities, network stability, and response speed [22]. At the current rate of data growth, the cloud’s computing power becomes less and less proficient in meeting these requirements. However, uploading all the data to the cloud will cause serious network congestion. Due to the limited network bandwidth, the data generated by a large number of IoT devices will impose a lot of pressure on the network bandwidth, causing cloud computing to no longer meet the requirements of latency and response speed in these scenarios. In addition, these data may have a large proportion of noise and errors. Some survey shows that only one third of the data obtained by most sensors are correct [23]. Putting these worthless data into the cloud will cause a huge waste of cloud server resources and a waste of network bandwidth.

Cloud computing has outsourcing features. Users need to host local data to the cloud when using cloud computing. This leads to a series of data security and privacy issues [21]. The data loss during long-distance transmission between devices and the cloud can damage the integrity and accuracy of the data. In addition, highly centralized computing and storage can also become serious problems. When one device in a centralized system goes wrong due to benign errors or malicious attacks, other devices will be negatively affected. The data privacy problem refers to the theft and utilization by other unauthorized persons, companies or organizations. Actually, data owners have lost control of their data uploaded to the cloud, so it is difficult to guarantee data privacy [24].

Computation offloading was originally proposed in cloud computing. The definition is that the terminal devices with limited computing power delegates part or all of the computing tasks to the cloud for execution. Similarly, computing offloading in EC refers to the problem that terminal devices with limited computing power delegate part or all of its computing tasks to the edge [30]. The main considerations are whether terminal devices will offload, how much they will offload and to which nodes they will offload. Computing offloading solves the problems of insufficient resources and high energy consumption in terminal devices.

Traditional methods of computing offloading applied to cloud computing are based on many assumptions, including that the default server has sufficient computing power and does not care about its energy consumption or network condition. However, traditional methods based on the above assumptions are not suitable for solving the computing offloading in EC where edge devices and servers have limited computing capabilities [31]. Reasonable computing offloading strategies are able to reduce energy consumption and latency. Therefore, computing offloading is an important research topic for optimizing EC.

Compared to traditional cloud computing, the most prominent advantage of EC is that it does not need to upload all the data to the cloud for computing and storage tasks, which largely frees up network bandwidth and other resources occupied by cloud computing. In the meanwhile, since tasks are distributed on each edge node with limited resources, an intelligent and efficient solution for resource management is crucial for EC.

EC also faces new challenges regarding data security and privacy [32]. Some of these challenges come from the inherent problems of cloud computing, and others come from the distributed and heterogeneity nature of EC itself [33]. Traditional solutions for data security and privacy issues of cloud computing are not applicable to the non-centralized computing model of EC. Therefore, further improving data security and further protecting data privacy is a problem worthy of researchers’ attention.

In detail, EC brings benefits to the application of AI. With the advent of the big data era, the widespread application of AI in people’s daily lives has become an irresistible trend. Of course, this trend still faces challenges. For example, AI’s reasoning and training requires strong computing power and sufficient energy support, but terminal devices often do not meet these two requirements. In recent years, cloud computing has fulfilled these needs by offloading AI model training and reasoning tasks that terminal devices cannot perform to the cloud server. However, relying solely on cloud computing will cause problems like insufficient bandwidth and high latency when a large number of AI models are used by a large number of terminal devices [48]. With the advent of EC, AI can be deployed near terminal devices and users on the edge and terminal with certain computing resources and storage resources, therefore meeting the needs for low latency and high network stability [11].

In return, EC also brings three ideas to the application of AI in other fields (visually represented by Figure 4).

As can be seen from the above, these three ideas have their own advantages and disadvantages, so existing studies are more inclined to choose the best idea according to the specific situation.

AI is playing an important role in the optimization of EC [52]. Since EC is distributed and the workload of each edge device changes dynamically with time and location, this uncertainty and unpredictability have brought huge obstacles to the application of EC. In this sense, EC still needs to be optimized and improved in many aspects, such as optimizing computing offloading, optimizing resource allocation, reducing latency and energy consumption, and improving user experience.

Many optimization problems in EC are very complex non-convex problems. As the number of devices and users increases, the scale of these problems will also rapidly increase [53]. Compared to traditional methods, ML is more suitable for solving optimization problems of EC and has better results [54]. In addition, AI algorithms are also good at effectively mining hidden information and laws from data in complex and noisy EC environments, which has plagued traditional optimization methods for a long time.

The traditional ML algorithms in this work particularly refer to those ML algorithms other than DL and RL. Given the availability of label information, the traditional ML algorithms can be divided into supervised learning, semi-supervised learning, and unsupervised learning. Among them, supervised learning requires labeled data to train the model, while unsupervised learning can autonomously discover the principles implicit in the data. As a hybrid of supervised learning and unsupervised learning, semi-supervised learning has access to both labeled data and unlabeled data. For example, the common supervised learning methods include support vector machines (SVM), boosting, and random forests; the common semi-supervised learning methods include label propagation and graphical models; the common unsupervised learning methods include clustering algorithms such as K-means and dimension reduction algorithms such as principal component analysis (PCA).

There are some obvious shortcomings of traditional ML algorithms. For instance, they are sensitive to data sets, the data become less effective when the data set is large enough, and they need complicated artificial feature engineering. In spite of these shortcomings, traditional ML has small energy consumption, small computing power cost, and is easy to deploy compared to DL and RL. Due to the distributed nature of EC, the appropriate AI algorithm can be reasonably selected according to the resource situation and task requirements of each edge and terminal device, so traditional ML can also rely on these advantages to find its place in EC [55].

DL resembles the functions of human brains. It has the ability to autonomously learn high-level features from raw data, thereby efficiently performing classification and prediction tasks [56, 57]. DL is usually deployed in a multi-layer structure. These layers can be fully connected layers, convolutional layers, pooling layers, normalization layers, or activation layers. A DL algorithm can be formed by the free combination of these layers. The more layers the algorithm includes, the “deeper” it is. The input of a neuron in each layer is the weighted sum of the outputs of the neurons in the previous layer. After the input is activated by an activation function, the obtained number is used as the output of the neuron [58]. Compared to traditional ML algorithms, DL has a more powerful ability to extract high-level features from massive data due to its multilayer structure [59].

The common DL models include: deep neural networks (DNN), convolutional neural networks (CNN), recurrent neural networks (RNN), and so on.

When a large number of labeled data are available, compared with traditional ML algorithms, DL performs better in natural language processing, computer vision and many other fields [57]. The characteristics of EC make the data collected from the physical environment can be processed locally, which meets the requirements of DL. Therefore, some EC studies also focus on using DL in EC anomaly detection [66], task scheduling and resource allocation in EC [67], and privacy protection [68].

Unlike supervised learning and unsupervised learning that rely on static data, RL is a learning algorithm that trains models through dynamic interaction with the environment. The core idea is that agents receive the state of environment and make actions to maximize the reward according to historical experience. Because reinforcement learning is good at solving decision-making problems, some studies [69, 70] have adopted RL algorithm in the decision-making of EC resource management, allocation, and scheduling.

Typical algorithms in RL are model-free and value-based Q-learning algorithm [71]. Each iteration of Q-learning algorithm will calculate an expected cumulative reward, called the Q-value, according to current state and given action. However, as the environment becomes more complex, the state space and action space will expand exponentially, thus reducing the convergence speed and taking up a lot of memory [72].

To solve this problem, deep Q network (DQN) [73] is proposed, which utilizes a DNN to approximate the Q-values. Compared with the classical RL algorithms, DQN has three advantages in dealing with EC with high complexity [74]. First, it is able to deal with high dimensional and complex systems. Second, it can learn the regularity of system environment. Last but not least, it is able to make optimal decisions based on current and past long-term reward. Therefore, some studies [75, 76] use DQN algorithms to optimize the control decision-making problems in EC and obtain good results.

However, DQN also has its shortcomings. Especially, when using nonlinear functions such as neural network to approximate the Q-function, the learning result of DRL is unstable or even divergent. To solve this problem, an experience replay mechanism using the prior experience is integrated into DQN [77, 78].

Federated learning (FL) is a distributed ML framework, which can effectively help multiple organizations train models under the requirements of user privacy protection, data security, and government regulations [79]. In this framework, different local users do not need to put all the raw data on the central server for training, but train the local model through privacy related data, then all the local models are aggregated into a global model on the central server [80].

As discussed above, the goal of EC is to deploy computing tasks at the edge of the network near the client. However, the data of a single edge node may not meet the requirements of model training. Therefore, the cooperation model training between different nodes under data privacy protection is a research hotspot; see, e.g., Reference [81].

Evolutionary algorithms are a kind of optimization methods inspired by biological evolution mechanism and biological behavior [82]. Evolutionary algorithms include particle swarm optimization (PSO), genetic algorithm (GA), differential evolution (DE), and so on.

Generally speaking, evolutionary algorithms are divided into the following steps. The first step is to initialize variables. After that, the evolutionary algorithms continuously iterate three steps named fitness evaluation and selection, population reproduction and variation, and population updating [82]. Finally, the second step is iterated until the termination condition is satisfied.

At present, evolutionary algorithm has been applied in many problems of EC, such as resource scheduling optimization [83], load balancing [84], and task scheduling [85]. In this article, we mainly discuss ML, a recently popular AI subclass, so evolutionary algorithm is only briefly introduced here.

At present, more and more studies have begun to make full use of AI to solve computing offloading [86]. We will summarize the AI-based computing offloading schemes in existing research to reduce energy consumption, reduce latency, and reduce both.

Reducing energy consumption. In terms of reducing energy consumption, a partial computing offloading scheme based on DL decision-making is proposed by Ali et al. [31]. The authors establish a new type of decision-making process, which can intelligently select the optimal computing offloading strategy, thus reducing the total energy consumed in the execution of computing tasks. Compared with its previous work in Reference [87], this strategy additionally considers the energy consumption of user equipment in the cost function, which reduces its energy consumption by 3%.

Reducing latency. Although EC itself has the advantage of low latency compared to cloud computing, it still has room for optimization. Smart-Edge-CoCaCo [88] is proposed to minimize the latency by jointly optimizing the wireless communication model, the collaborative filter caching model, and the computing offloading model. In addition, since the computing power of edge devices is limited, offloading all tasks to edge devices may exceed the capacity of the edge device. With this in mind, Xu et al. [89] propose a DL-based heuristic offloading method. This method uses origin-destination electronic communications network distance estimation and heuristic searching to find the optimal computing offloading strategy.

Reducing both energy consumption and latency. All the methods mentioned in previous paragraphs either only minimize energy consumption, or only minimize latency. There are also studies that consider the minimization of both through RL. Kiran et al. [54] propose a scheme that uses Q-learning to make optimal control decisions to reduce the delay in EC and adds constraints to the cost function to reduce energy consumption in EC. Although this scheme has a good effect on reducing energy consumption and delay, it does not take into account the curse-of-dimensionality problem of EC.

The curse-of-dimensionality refers to the problem that the complexity of the problem solving will increase at an exponential speed as the dimensionality increases [90, 91]. To solve the curse-of-dimensionality problem, Xu et al. [91] propose an algorithm that uses the special structure of state transitions of the considered EC system to overcome the curse-of-dimensionality problem. It is worth noting that the authors use energy harvesting [92] to reduce the consumption of traditional energy by fully utilizing renewable energy, but the transmission delay model and the energy consumption model are required to be known (this requirement can be eliminated by the method proposed in Reference [93]).

Compared with RL algorithms, DRL algorithms have stronger abilities to deal with high-dimensional state space. Therefore, Cheng et al. [94] propose a model-free DRL-based computing offloading method based on a space-air-ground integrated network to reduce EC latency and energy consumption. This method uses Markov decision process to represent the computing offloading decision process, and uses DRL to learn network dynamics.

Yet the ability of DRL algorithms to cope with high-dimensional state space is not perfect in every respect. Chen et al. [95] propose a new DNN model based on function approximator, and they also adopt double deep Q-network so that the optimal offloading strategy can be discovered without prior knowledge. Similarly, Lei et al. [10] propose a new type of value function approximator to deal with high-dimensional state equations. The authors also use an infinite-horizon average-reward continuous-time Markov decision process to represent the optimal problem. Finally, DRL is applied to solve the optimal computing offloading decision to reduce the energy consumption and latency of EC.

The DRL-based methods mentioned above use a centralized style for model learning. However, there is a potential assumption in this style that edge devices in EC have sufficient computing power. In fact, many edge devices do not yet have such powerful computing capabilities. As a result, Ren et al. propose a distributed computing offloading strategy combining federated learning and multiple DRLs [96]. It is proved by experiments that this method outperforms the centralized learning method in reducing the transmission cost in EC. In addition, distributed learning also has the advantage of fast convergence [97]. This is proved in Reference [98] by the method of optimizing computing offloading through distributed ML.

EC provides certain computing capabilities near the data source, so that many computing tasks do not need to be delivered to the cloud for execution. While this model brings high response speed to people, it will inevitably cause a surge in energy consumption on the edge side. Moreover, many applications in EC require AI algorithms to make real-time decisions (such as intelligent driving [99] and intelligent monitoring systems [100]), but AI algorithms are computationally intensive to varying degrees. This is a huge challenge for devices with limited power. From the perspective of overall energy consumption, with the gradual popularization and widespread application of AI, how to control global overall energy consumption or improve energy efficiency is also very important.

Apart from computation offloading, there are many other factors that affect the energy consumption of edge devices. For example, different AI algorithms and different hardware structures adopted by edge devices will also affect energy consumption [101]. We will introduce AI solutions to reduce EC energy consumption in terms of optimizing hardware structure, controlling operating status, and combining energy Internet.

Optimizing hardware structure. A static random access memory (SRAM) [102] is able to reduce memory data throughput, and it combines parallel CNNs to enable simultaneous access to different memory blocks. Experiments show that this architecture significantly reduces energy consumption compared to traditional digital accelerator using small bitwidths. Based on field-programmable gate array (FPGA) [103], Lammie et al. [104] design a binarized DNN accelerator for weed species classification, which reduces energy consumption by 7 times compared with GPU-based accelerator under the same conditions. The authors believe that well-cultivated FPGA-based accelerator for AI algorithms is an ideal choice for edge devices with limited resources but need to perform learning and reasoning tasks.

Controlling operating status. Sun et al. propose a method based on DRL to reduce the medium and long-term energy consumption of EC by controlling the communication modes of user devices and the light-on state of processors [105]. This method uses Markov process to model the energy consumption of cache states and cloud processors and DRL to make decisions. According to some constraints (quality of service constraints, transmission power constraints, and the computing capability constraint in the cloud), the method uses an iterative algorithm to optimize the precoding of user devices.

Combining Energy Internet. EC has distributed characteristics, and the workload of edge-side devices will dynamically change with different geographical locations and times, which makes the energy consumption of each edge node unpredictable and uneven. To deal with the huge energy demand of EC and its heterogeneity, the combination of energy Internet (including smart grid and microgrid) with EC can provide renewable energy for EC [70, 106]. Energy Internet is a distributed energy production model that achieves local energy self-sufficiency by making full use of renewable energy sources [107, 108]. This feature of energy Internet is very suitable for providing energy to EC, thereby reducing the consumption of non-renewable energy. Since renewable energy is infinite, reducing non-renewable energy consumption is also equivalent to reducing energy consumption. However, due to the uncertainty of renewable energy production [109], some studies [70, 106] also aim to balance the energy supply and demand of EC through DRL-based control strategies. With the deployment of EC devices into energy Internet, energy management will also become more complex [110]. DRL combined with curriculum learning [111] has been used to realize a bottom-up energy management scheme [110].

Delegating computing and storage tasks from the cloud to the edge can reduce the security problems caused by network congestion and centralization to some extent. However, the distributed environment of EC also brings new security problems, such as distributed denial of service (DDoS) attacks and jamming attacks that cause illegal distribution of distributed system resources [33, 112]. What was previously applicable to a centralized environment (like cloud computing) is no longer applicable to solving these new security issues. In this part, we will review the studies on improving the security of EC based on AI algorithms.

Traditional machine learning methods. Traditional ML can help with the identification and classification of different attacks. In response to jamming attacks that threaten EC security, Wang et al. [113] propose a stochastic game framework that maximizes the spectral efficiency throughput by minimax-Q learning, thereby gradually learning the optimal strategy. The disadvantage of this method is that it needs extra bandwidth to avoid jamming attacks. This can be avoided by selecting the most reliable server based on online learning to reduce the security risks caused by jamming attacks [114]. To reduce the false alarm rate and data transmission delay of traditional intrusion detection systems, an algorithm selection mechanism can be deployed on the edge side [115]. This enables intelligent selection of the optimal ML algorithm for edge devices to distinguish false alarms. The experimental results prove that the method based on AI algorithm can improve the security of EC more effectively than the method based on non-AI algorithm.

Among various network attacks, DDoS is a relatively common attack method. Hypergraph clustering [116] can be adopted to model the relationship between edge nodes and DDoS to improve the recognition rate [117]. Kozik et al. uses a single-layer neural network to build the extreme learning machine classifier [112]. In this method, the training task of the attack detection classifier model is performed in the cloud with powerful computing resources. The trained classifier model is then offloaded to the edge devices for attack detection. In addition, experiments have also proven that the extreme learning machine classifier has faster convergence speed and stronger generalization performance than most traditional classification algorithms (such as SVM, or single-layer perceptron).

DL methods. Although traditional ML algorithms can improve the accuracy and robustness of network attack detection and recognition, they lack the ability of automatic feature extraction [118]. As a result, traditional AI algorithms are not sensitive to known but slightly changed attacks. At the same time, due to the lack of prior knowledge of unknown vulnerabilities, they can not effectively detect zero-day attacks [119]. Deep learning, however, has been successfully applied in image processing, computer vision and many other fields in recent years because of its structure that can automatically mine and learn the hidden features in massive data [63]. Researchers begin to focus on DL, since the problem of cyber-security attack identification in EC is similar to the tasks in these fields.

Abeshu et al. [56] propose a DL-based method for attack detection in EC. To reduce the burden of model training and improve the accuracy of the model, this method uses a pretrained stacked autoencoder to screen the real valuable features and then uses softmax to do classification. This method shows great advantages in the aspects of availability, scalability and effectiveness compared with traditional ML algorithms. However, the authors fail to take into account the improvement of the detection rate of new attacks. This can be solved by unsupervised learning. The DL-based algorithm proposed in Reference [120] learns the characteristics of the attack through the deep belief network and uses the softmax function to identify various attacks on the EC. The difference is that this solution incorporates unsupervised learning restricted Boltzmann machines into the proposed model. Since unsupervised learning restricted Boltzmann machines is a stochastic artificial neural network with active learning characteristics, this model enables active learning to improve the recognition rate of attacks that have never occurred before.

To a certain extent, EC reduces the risk of privacy leakage caused by uploading data to cloud servers that users cannot control. However, the problem of data privacy leakage also exists on the edge side. On the one hand, the distributed nature of EC brings new challenges to privacy protection. On the other hand, the application of AI on the edge side requires massive data for model training and reasoning, which are inevitably mixed with a large amount of user privacy. During the training process, some models may save part of the training set with private data, so an attacker can illegally obtain users’ privacy by analyzing these models [121]. Consequently, it is very important to ensure the data privacy and security of edge-side users without affecting the performance of EC. This topic has attracted the attention of many researchers in recent years.

Post-decision state learning. A post-decision state (PDS) learning method is proposed in Reference [122], in which the state transition function is factored into known and unknown components. This method first uses the Markov decision process to describe EC’s offloading problem and then solves the problem by combining PDS-learning technique with the traditional deep Q-network algorithm. This combination can well balance task scheduling and privacy protection. It is worth noting that compared with the traditional deep Q-network, the new algorithm can speed up the model training by learning some additional information (such as the energy utilization of edge devices).

Federated learning. A privacy-preserving asynchronous FL mechanism (PAFLM) for EC is proposed, which allows multiple edge nodes to realize more efficient FL without sharing private data and affecting inference accuracy [81]. Because the local model training of each node depends on the data inside the node to a large extent, it is easier to lead to local optimum. Through FL, the local model can be optimized with the help of the model parameters of other nodes, which can solve local optimum problem and improve the accuracy of model.

Differential privacy. To protect the user privacy in the training data set under EC, AI algorithms are usually combined with differential privacy, a system where including or excluding any piece of data will not change the results of related data analysis to a great extent [123]. In other words, by applying differential privacy, observers cannot tell from its output if any particular piece of information has been used [123]. Du et al. [124] propose two AI-based algorithms that satisfy differential privacy: objective perturbation algorithm and output perturbation algorithm. The difference between the two is that objective perturbation adds Laplace noise to objective functions, while output perturbation adds the noise to outputs. By injecting Laplace noise, ML algorithms show better efficiency and accuracy in prediction, and they are more effective in protecting the privacy of training data used in EC. Similarly, a deep reasoning framework based on differential privacy, called EdgeSanitizer, is proposed in Reference [125]. The framework uses as much useful information as possible with a DL-based data minimization method. Then it removes as much sensitive private information as possible from data sets by adding random noise to the original data through a local differential privacy method [126]. This approach ensures that the data is used to the maximum extent while protecting the privacy in EC.

DRL has been proven to be capable of handling dynamic decision problems with high-dimensional states and action spaces [127]. At present, some studies have focused on DRL to solve the resource allocation problem in EC.

The method in Reference [77] captures the fact that the EC environment state is constantly changing. The information about wireless channel conditions, each node’s trust value, the contents in the cache, and the vacant computational capacity is passed to the DNN to estimate the Q-function. The network operator’s revenue is regarded as the reward, and the agent trains the DNN through the obtained reward. It avoids local convergence by adjusting the learning rate. Although this method has a good effect, there is still room for improvement in convergence and performance.

Although the study above proves that DQN has a good performance in optimizing dynamic decision problems with high-dimensional state space, there are still some limitations when solving problems based on high-dimensional action space. Therefore, Chen et al. [127] propose a new DRL-based resource allocation decision framework that makes the following two contributions:

The experiment proves that compared with the method of directly using DQN, this method has reduced the delay by 51.71%.

Smart cities need to continuously monitor the infrastructure and operation of the city, and they need to make quick judgments and respond quickly to security incidents. Integrating AI algorithms can improve the accuracy of security event identification. However, the network bandwidth is limited, and excessive data transmission will cause instability in network transmission. How to deal with massive data is therefore a very difficult problem for real-time monitoring systems. EC performs most of the data processing and analysis tasks on the edge and transmits only part of the data to the cloud. This can greatly reduce the network transmission pressure caused by massive monitoring data while improving the response speed of the application.

To ensure the safety of urban residents in public places or private places, a series of monitoring systems (e.g., traffic monitoring, indoor and outdoor monitoring, facility monitoring, violence and crime detection) need to be widely deployed to analyze and tackle the surrounding environment in real time. In urban monitoring, for instance, person re-identification is an important part to ensure the safety of residents. A new Siamese network architecture for person re-identification is proposed in Reference [131]. This architecture speeds up the retrieval of pedestrians by introducing EC. Considering that traditional methods may learn poorly and inefficiently due to the low resolution of images, together with the limited computing power on the edge side, the architecture introduces a residual model layer that can mine deep features and reduce the complexity of the global average pooling layer.

Utilizing the distributed characteristics of EC and the geo-distribution characteristics of monitoring data, it is a good idea to apply different AI algorithms to EC in a distributed way. A monitoring system based on distributed deep learning model is mentioned in Reference [100]. By introducing EC, the system reduces the cost of communication and improves response speed. This article uses the distributed characteristics of the edge side to deploy a distributed DL training method based on task-level and model-level parallel training. The goal is to speed up the training of the sub-model by taking advantage of different learning models while also using the computing power of edge nodes.

In contrast, Tang et al. [132] adopt the idea of configuring different AI algorithms in the edge and the cloud. The proposed general-purpose EC architecture for urban pipeline monitoring systems takes advantage of the low latency of edge nodes so that pipeline faults can be discovered in time, and response decisions can be made quickly. The architecture consists of four layers, and the architecture deploys different AI algorithms and control strategies in different layers to achieve low latency, low energy consumption, and high accuracy for smart pipeline monitoring to ensure the safety of pipelines in cities.

Challenges. In the process of protecting urban security, data privacy and security are also crucial. AI is an effective method of identifying malicious attacks and preventing privacy leakage, but the computing resources of edge devices are limited. Therefore, it is still a major challenge to design lightweight and effective AI algorithms suitable for EC [131].

With the popularity of IoT and cloud computing, more and more personal medical devices are being used in daily life. These devices can collect users’ physical data and upload the data to a cloud server. Through AI analysis, these data can greatly improve the accuracy of medical systems for disease classification and diagnosis. However, this model of cloud computing cannot really meet the requirements of telemedicine for time delay and data transmission.

Compared with traditional cloud computing, the application of EC meets the requirements of medical system for stable data transmission, transmission delay, and data security. In some emergency situations, for example, just the occurrence of errors such as long response time or data loss may directly threaten human life. Besides, EC has strong location awareness characteristics [33]. The higher processing speed of EC becomes a critical factor for location-sensitive medical systems.

Next, we will summarize existing urban medical and residents’ health works that use EC to improve AI algorithms in terms of remote diagnosis and early warning of diseases, infectious disease prevention and control, and smart assessment.

Remote diagnosis and early warning. Muhammad et al. [133] propose a voice disorder assessment and treatment system. The sound data collected by the system is pre-processed by edge devices before being uploaded to the cloud. The system configures the CNN model to the edge server, so that the edge side has the capability of voice disorder detection and classification. Compared with the method without EC architecture in Reference [134], this method has lower latency and can effectively reduce the pressure on network bandwidth. However, this system still needs to send the diagnosis to a human expert, and the human expert decides the treatment plan.

For some diseases that are not easy to detect at an early stage and those that can be best treated in the early stages of the disease (e.g., lung cancer), the patient’s survival can be significantly extended if a patient is diagnosed and treated early in the disease [135]. To improve the early diagnosis rate and accuracy of lung cancer, a lung cancer diagnosis system based on EC and AI is proposed in Reference [135]. This system can not only improve the early accuracy of lung cancer but also improve the efficiency and security of diagnosis. In the future, how to combine EC and AI algorithms to diagnose diseases and generate corresponding treatment plans without a human doctor is a valuable research direction.

Infectious disease prevention and control. The use of EC’s powerful location awareness feature can effectively strengthen the prevention and control of infectious diseases. The healthcare framework proposed in Reference [51] can diagnose whether a user has been infected by Kyasanur forest disease and can map out areas where infectious diseases are likely to occur on the map. The network edge near the data source in this structure is responsible for data preprocessing, model training and reasoning. To more accurately identify infected people and outbreak-prone areas, this layer incorporates a classifier called EO-NN, which combines hybridization of the extremal optimization (EO) and the neural networks (NN). Once a new infected person is detected, it will inform the infected person and nearby hospitals immediately. With the distributed nature of EC, the system has the ability to identify areas prone to infectious diseases.

Smart assessment. Residents’ daily dietary structure management is also an important part of urban medical care, which also plays an important role in the prevention of diseases. Based on food image recognition, Liu et al. [49] propose a dietary assessment system under an EC architecture. The edge layer between end users and the cloud can minimize the response time and energy consumption, and the CNN algorithm can improve the accuracy of recognition. Compared to the previous system in Reference [136], which is only suitable for small data computing tasks, this system has the ability to perform large-scale data computing tasks.

Challenges. Medical diagnosis needs accurate judgment, which requires AI algorithms to extract all useful information from big data. However, the useful information that can be obtained by existing algorithms is rather limited. For supervised learning, manual labeling of data may also lead to unknown mistakes. In addition, the data acquisition system of smart medical in the future will be mainly deployed on wearable devices. To quickly analyze and respond to the collected data, it is also an important direction to deploy AI model to these wearable devices [136], which poses a great challenge to the energy supply of devices. How to balance the accuracy and lightweight of AI models is a direction worthy of studying [137].

The trend of urbanization is also prompting the rapid increase of energy consumption in cities. This poses many challenges for urban energy management. For example, to meet the city’s demand for energy, energy companies need to produce excess electricity to ensure continuous energy supply to the city. This leads to a certain degree of waste of energy [138]. In the era of big data, a large number of sensors deployed in various corners of the city can obtain data related to energy consumption in real time. These data include population density, electricity usage, and a wealth of environmental information that helps predict energy consumption and energy management. In addition, applying AI algorithm to energy management has greater advantages than traditional methods [139]. Under these conditions, the introduction of EC and AI can make energy consumption prediction and energy management faster and more accurate. A typical EC-based smart city energy management architecture is shown in Figure 5.

Real-time energy management decisions require dynamic predictions of energy consumption. However, the complexity and diversity of energy data and the dynamic nature of IoT data make it rather difficult to build an effective energy prediction system. In response to this problem, Liu et al. [140] design an EC-based energy management framework for reducing energy consumption in cities. Under this framework, the authors propose two DRL-based energy scheduling strategies:

The authors prove by experiment that cloud-edge collaboration works best in terms of energy consumption, followed by the method of deploying AI algorithms only on the edge side, and the worst is the method of deploying AI algorithms only on the cloud [138]. This also indicates that EC is not a substitute for cloud computing, and the relationship between the two should be synergistic and complementary.

Challenges. The rapid growth of the number of edge devices deployed to cities has exacerbated the global energy crisis and global warming. One way to alleviate this problem is to use renewable energy to power edge devices. Considering that edge devices are scattered in different locations of the city, the energy consumption of traditional energy can be greatly reduced by using distributed renewable energy generation devices. However, this solution still faces many challenges, such as how to minimize the consumption of traditional energy while ensuring the normal operation of edge devices, and how to establish a complementary power system for different edge devices [140]. As a control center in EI system, energy router needs certain computing power [141, 142]. Therefore, it is also a feasible idea to combine energy router with EC in future research.

The rapidly changing network structure, communication status, and computing load have led to the dynamics and uncertainty of task offloading [150], making efficient task offloading and resource allocation decisions more difficult. Feng et al. [148] use the ant colony optimization algorithm with fast convergence to solve the NP-hard task assignment problem. This method establishes multiple objective functions, and uses heuristics algorithm for optimization. However, this method is not good at making optimal decisions for offloading multiple data dependency tasks. In response to this problem, an EC framework for obtaining the optimal solution of task offloading through DRL is proposed in Reference [149]. The framework takes into account data dependencies, as well as resource requirements, vehicle movements, and access networks. It uses the asynchronous advantage actor-critic (A3C) algorithm [151] for the online optimization of task offloading decision to adapt to the dynamic changes of the vehicular network. Edge nodes will first distribute the trained decision model to the surrounding vehicles, and then upload the decision model online after vehicles’ complete learning. To improve the performance of resource allocation and management, the prediction of wireless channel parameters is a very important means. Liu et al. [152] use LSTM to excel in spatio-temporal correlation in channel parameters and propose a wireless channel parameter prediction model based on LSTM and EC to optimize resource allocation and task scheduling in vehicular network.

In IoV, energy consumption is a huge obstacle that restricts its development. However, the studies mentioned above fail to consider the issue of energy consumption while making optimal offloading decisions. Yang et al. [53] put forward a joint optimization problem consisting of power control, user association, and resource allocation to minimize energy consumption in IoV. Finally, the feasible solution of this problem is obtained by an algorithm based on fuzzy c-means clustering that allows one data point to join multiple clusters.

The maturity and application of autonomous driving technology will bring more free time to passengers and drivers in the future. This will increase passengers and drivers’ demand for on-board entertainment, such as listening to music, watching videos, and more [153]. These on-board entertainment activities have extremely high requirements for network latency, so implementing these computing-intensive applications in a connected vehicle with limited resources is facing great challenges [154]. These challenges include how to efficiently cache network content and how to efficiently schedule tasks and allocate resources.

The traditional content caching method is to cache the current popular content in roadside units in advance, but this also causes a waste of storage resources. To coordinate passenger experience and content caching costs, Hou et al. [153] propose a Q-learning-based caching strategy under the EC architecture. The action of this caching strategy consists of two parts, one is the cache amount, and the other is the roadside units to which the content is cached. The reward of this caching strategy is the elapsed time of transmitting the content required by the user. In addition, this article uses LSTM to predict the driving direction of the vehicle to better select roadside units.

In contrast, the method of Reference [155] imposes the task of content caching on both roadside units and vehicles. It uses a collaborative model based on Q-learning vehicles and roadside units for content caching and computation distribution. This model can make full use of the limited storage and computing resources of vehicles. In other words, the system will select vehicles and roadside units to perform the tasks of caching and computing according to the position and direction of motion of the car requesting the service. If the vehicles and roadside units around the car cannot meet their requirements, then the cache and calculation tasks will be handed over to the base station.

Aiming at the challenges of executing compute-intensive applications on cars with limited resources, Ning et al. [154] first use finite-state Markov chains to model vehicle-to-infrastructure communication and computing states and then express the resource allocation and task scheduling strategy as a goal to maximize users’ quality of experience (QoE).

In addition to the macro-control of resource allocation, it is also an important research direction to give AI technology to vehicle intelligence under the EC architecture [156]. For example, Ferdowsi et al. [157] propose an EC architecture that integrates DL to handle complex vehicle and traffic information. The architecture enables functions such as vehicle automatic control and driving route analysis. This architecture uses different DL algorithms according to the characteristics of different problems:

The increasing number of vehicles aggravates the problem of traffic jam. Traffic scheduling is a very effective way to deal with this problem. However, due to the large number of vehicles and the scale of road network, the number of routes that vehicles can choose increases exponentially. Therefore, it is not feasible to use centralized controller for route planning. Based on this problem, a distributed cooperative routing algorithm based on evolutionary game theory is proposed in Reference [158]. Each edge node deploys a roadside unit (RSU), in which normal RSU is responsible for collecting traffic information, and game RSU controls nearby vehicles through proposed evolutionary game strategy.

The combination of EC and IoV improves the response speed of vehicle scheduling and control, which further promotes the vehicle intelligence. However, there are still some challenges [159]. For example, when the vehicle is moving at a high speed, its communication connection needs to be switched between different edge servers, which may lead to a series of problems, such as disconnection or the degradation of user experience. In addition, one of the cores of IoV systems is resource sharing between different vehicles. As a result, how to set a reasonable incentive mechanism to encourage participants to share resources is vital. Finally, resource sharing will also bring some data privacy and security issues [160].