Deep Learning For Edge Computing Applications A ST

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2020.2982411, IEEE Access
Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier xxx/ACCESS.2017.DOI
Deep Learning for Edge Computing

Applications: A State-of-the-art Survey
FANGXIN WANG1 (Student Member, IEEE), MIAO ZHANG1 , XIANGXIANG WANG1 ,
XIAOQIANG MA2 , AND JIANGCHUAN LIU1 (Fellow, IEEE)
1
School of Computing Science, Simon Fraser University, Burnaby, BC V5A 1S6, Canada
2
School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan 430074, China
Corresponding Author: Xiaoqiang Ma ([email protected]).
This work was supported in part by a Canada NSERC Discovery Grant and the National Natural Science Foundation of China under Grant
61702204.
ABSTRACT
With the booming development of Internet-of-Things (IoT) and communication technologies such as
5G, our future world is envisioned as an interconnected entity where billions of devices will provide
uninterrupted service to our daily lives and the industry. Meanwhile, these devices will generate massive
amounts of valuable data at the network edge, calling for not only instant data processing but also intelligent
data analysis in order to fully unleash the potential of the edge big data. Both the traditional cloud computing
and on-device computing cannot sufficiently address this problem due to the high latency and the limited
computation capacity, respectively. Fortunately, the emerging edge computing sheds a light on the issue by
pushing the data processing from the remote network core to the local network edge, remarkably reducing
the latency and improving the efficiency. Besides, the recent breakthroughs in deep learning have greatly
facilitated the data processing capacity, enabling a thrilling development of novel applications, such as video
surveillance and autonomous driving. The convergence of edge computing and deep learning is believed
to bring new possibilities to both interdisciplinary researches and industrial applications. In this article,
we provide a comprehensive survey of the latest efforts on the deep-learning-enabled edge computing
applications and particularly offer insights on how to leverage the deep learning advances to facilitate edge
applications from four domains, i.e., smart multimedia, smart transportation, smart city, and smart industry.
We also highlight the key research challenges and promising research directions therein. We believe this
survey will inspire more researches and contributions in this promising field.
INDEX TERMS Internet of Things, Edge Computing, Deep Learning, Intelligent Edge Applications.
I. INTRODUCTION the cloud servers will carry out the computation-intensive

With the explosive development of the Internet-of-Things tasks and distributed the result back. However, the large
(IoT) as well as the communication technologies such as latency caused by the long physical distance can sometimes
WiFi and 5G, our future world is envisioned as an inter- become unacceptable, especially for those latency-sensitive
connected entity where billions of digital devices would applications like autonomous driving and highly interactive
provide uninterrupted services to both our daily lives and the applications such as VR gaming. In addition, the huge data
industry. As reported by Cisco [1], there will be more than communication also greatly increases the pressure of the
50 billion IoT devices connected by the Internet by 2020. backbone network, bringing large overhead and cost to ser-
Such numerous IoT devices will generate a myriad of valu- vice providers.
able data which, once being well processed effectively and The emerging edge computing [2] provides a promising
efficiently, can empower many groundbreaking applications. solution for this problem. Though with many representa-
Traditional computing architecture relies on cloud computing tion forms, such as fog computing [3] and cloudlet [4],
to provide sufficient computation capacity and sustainable the basic idea of edge computing is that the computation
energy. In this system, IoT devices are responsible to collect capacity should be deployed close to the data source for
the data and deliver it to the remote powerful cloud, and data processing, rather than transmitting the data to places
VOLUME 4, 2016 1
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
Applications
Edge
Nodes
Edge Computing Deep Learning

FIGURE 1: The illustration of deep learning enabled edge computing applications.
with computation power. In this way, massive numbers of al. [11] presented their understanding on edge computing and
servers are deployed at the edge of the network and the tasks deep learning from five aspects, while they did not make a
at IoT end devices can be offloaded to the edge servers for comprehensive and in-depth overview from the perspective
instant processing. The paradigm of edge computing brings of applications. Similarly, Chen et al. [12] focused on mul-
many benefits compared to cloud computing. First, since tiple aspects in deep learning and edge computing, but only
data computing happens closer to the data source, the com- mentioned a general abstraction for those emerging applica-
munication latency can be largely reduced, facilitating the tions. Zhou et al. [13] mainly focused on the deep learning
development of latency-sensitive applications. Besides, local model training and inference with edge computing. There are
computation can better protect data privacy and application also a series of surveys for mobile edge computing [2], [14]–
security. Last but not least, data processing at the network [16] and deep learning [17], [18], respectively, while they
edge can effectively reduce traffic at the backbone network focused on either of them without a comprehensive review
so as to alleviate the network pressure. on the combination. Therefore, a complete survey on the
Deep learning has made remarkable breakthroughs in recurrent cutting-edge researches is required at this time to
cent years due to the powerful perception ability. It has been provide a comprehensive review on deep-learning-enabled
widely used in various fields such as computer vision [5] and edge computing applications and illuminate the potential
natural language processing [6]. Besides, its performance in future directions.
computer and chess games, e.g., Atari Games [7] and the To fulfill this gap, in this article, we focus on the con-
game of Go [8], even exceeds the best level of human players. fluence of edge computing and deep learning, and conduct
The confluence of edge computing and deep learning will an up-to-date literature review on the latest advances of
undoubtedly sheds a light on address the current challenges, leveraging deep learning to empower the edge computing
enabling more desirable applications. On one hand, the ap- applications, as illustrated in Fig. 1. We first provide a brief
plications of edge computing urgently need the powerful pro- overview of edge computing and deep learning on concepts,
cessing capabilities of deep learning to handle various com- advantages as well as representative technologies. We then
plicated scenarios, such as video analytics [9], transporta- summarize the deep-learning-enabled edge computing appli-
tion control [10], etc. On the other hand, edge computing cations into four representative domains, i.e., smart multi-
has provided specifically designed hardware foundations and media, smart transportation, smart city and smart industry,
platforms to better support deep learning running at the edge, which cover a series of crucial applications like video analyt-
e.g., the light-weighted Nvidia Jetson TX2 developing kit1 . ics, autonomous driving, intelligent traffic control, industrial
Though lots of pioneer efforts have been made towards deep- manufacturing, etc. At last, we discuss some key research
learning-enabled edge computing applications, this field is challenges and promising research directions to achieve sta-
still in the infant stage. ble, robust, and practical edge learning applications. Differ-
Several existing surveys have investigated the convergence ent from existing surveys, this article focused on the deep
of deep learning and edge computing in the literature. Han et learning enabled edge computing applications, presenting a
comprehensive review and highlighting the challenges and
1 https://developer.nvidia.com/embedded/jetson-tx2 opportunities.
2 VOLUME 4, 2016
The rest of this article is organized as follows. We present interconnected through the distributed fog nodes to provide
the basic paradigm understanding of edge computing and collaborative services [22]. The fog nodes are also mainly
its advantages in section II. We introduce some deep learn- designed to provide better support for the IoT devices. From
ing techniques in section III. The review of deep-learning- this perspective, compared to other similar edge computing
enabled edge applications is summarized in section IV. We paradigms, fog computing often stands in alignment with IoT
highlight the challenges and research directions in section V. and emphasizes more on the end side.
We at last conclude this article in section VI.
3) Mobile Edge Computing
II. EDGE COMPUTING OVERVIEW The paradigm of mobile edge computing was first standard-
The emerging edge computing in recent years has seen suc- ized by European Telecommunications Standards Institute
cessful development in various fields given its great poten- (ETSI) [23], which aims to provide sufficient computing
tial in reducing latency and saving cost. Different from the capacities within the radio access network (RAN). It envi-
cloud computing architecture, edge computing enables data sions that the computing capacities are placed at the end of
processing at the edge of the network. On one hand, data the cellular network, e.g., the wireless base stations. Since
computing is put closer to the data source, which greatly base stations are the important access gate for numerous
facilitates the development of delay-sensitive applications. IoT devices, mobile edge computing could provide direct
On the other hand, the network traffic is largely reduced service to the end devices through only one hop, bringing
since the local processing avoids much data transmission, great convenience for IoT data processing [16].
which remarkably saves the cost. In this section, we briefly
introduce some edge computing paradigms and highlight the B. ADVANTAGES OF EDGE COMPUTING
key advantages of edge computing. Compared to traditional cloud computing, edge computing
has many unique advantages, including low latency, energy
A. EDGE COMPUTING RELATED PARADIGMS saving, context-aware service, and privacy as well as security.
The key component in edge computing is the edge devices, We next summarize them as follows.
which are usually edge servers located closer at the network
end for data processing, communication, caching, etc. There 1) Low Latency
are also some other paradigms or technologies that share Since edge devices are placed closer to end devices, which
similar concepts with edge computing. We next discuss and are usually both the data source and the transmission tar-
differentiate some typical paradigms that are related to edge get of processing results, the transmission latency can be
computing. largely reduced compared to the cloud computing scenario.
For example, the transmission latency is usually tens (or
1) Cloudlet hundreds) of milliseconds between an end user and a cloud
Cloudlet, initiated by Carnegie Mellon University, is en- server, while this number is usually several milliseconds or
visioned as small clusters with certain computation and even at microsecond level. The emerging 5G technology
storage capabilities deployed near the mobile devices such further enhances the advances of edge computing from the
as buildings and shopping centers for assisted processing, perspective of low latency transmission, which empowers a
offloading, caching, etc. Cloudlet usually utilizes virtualiza- series of emerging applications, such as autonomous driv-
tion management technologies [4] to better support mobile ing [24], virtual reality/augmented reality and healthcare-
applications. And an important target of cloudlet is to bring related applications.
the cloud advances to mobile users [19], achieving more
low-latency and resourceful processing. Micro data centers 2) Energy Saving
(MDCs) [20], initiated by Microsoft that are similar to the Restricted by the size and usage scenarios, IoT devices
concept of cloudlet, are a small-scaled version of data centers usually have quite limited energy supply, but they are also
to extend the hyperspace cloud data centers. Different MDCs expected to perform very complex tasks that are usually
are connected by the backbone network to achieve more effi- power consuming. It is challenging to design a cost-efficient
cient and intelligent computation, caching, and management. solution to well power the numerous distributed IoT devices
MDC also serves as an important role in managing numerous given that frequent battery charging/discharging is impracti-
Internet of Things (IoT) devices [21]. cal in not possible [16]. Edge computing enables the billions
of IoT devices to offload the most power-consuming compu-
2) Fog Computing tation tasks to the edge servers, which not only greatly reduce
Fog computing [3], first proposed by Cisco, is a computing the power consumption but also improves the processing
paradigm that aims to bring cloud computing services to the efficiency.
end of the enterprise network. In fog computing, the data
processing is carried out at fog nodes, which are usually de- 3) Context-Aware Service
ployed at the network gateway. The fog computing presents Context-aware computing [25] is playing an important role in
a high-level platform that the numerous IoT devices can be IoT and edge computing applications, since good modeling
VOLUME 4, 2016 3
Network Network
Policy Policy
Action
State Environment Action
State E
observe state
observe state
(a) Restricted Boltzmann Machine (b) Autoencoder (c) Convolutional Neural Network
Network Policy
Network Policy
Action
Action
State State Environment
Environment
observe state
observe state
(d) Deep Neural Network (e) Recurrent Neural Network (f) Deep Reinforcement Learning
Network Policy
FIGURE 2: The structures of different deep learning models.Network
Policy
Action
State Environment
Action
and reasoning of collected data can highly rely on the context work (CNN),State
recurrent neural network (RNN), andEnvironment
deep
of the data. With the advantage of the proximity nature, reinforcement learning (DRL). The basic architectures of
edge servers can collect more context information to support observe state
these learning models are illustrated in Fig. 2.
the data processing. For example, in the Amazon Go super-
observe state
market, video cameras can not only record the goods that A. RESTRICTED BOLTZMANN MACHINE
customers select but also predict customers’ interest based Restricted Boltzmann machine (RBM) is a kind of proba-
on their staying location, duration and behaviors. bilistic graphical models that can be interpreted as stochastic
neural networks [28]. A typical two-layer RBM includes a
4) Privacy and Security visible layer that contains the input we know and a hid-
Compared to cloud computing, edge computing is more den layer that contains the latent variables, as described in
efficient and effective in protecting the data privacy and Fig. 2(a). RBMs are organized as a bipartite graph, where
application security of users. On one hand, edge servers each visible neuron is connected to all hidden neurons
are usually geographically distributed clusters that could and vice versa, but any two units are not connected in
be managed and maintained by users themselves. Sensitive the same layer. RBMs have seen successful applications in
information can be monitored and protected more strictly. many fields, such as collaborative filtering [29] and network
On the other hand, the small-scale nature makes it more anomaly detection [30]. Multiple stacked RBM layers can
concealed than large-scale data centers, further making it less form a deep belief network (DBN), which consists of a
likely to become a target of attacks [26]. visible layer and multiple hidden layers. The training of a
DBN follows a layer-by-layer method, where each layer is
III. DEEP LEARNING METHODS treated as an RBM trained on top of the previously trained
Deep learning has been widely applied in many fields with layer [31]. Many applications can benefit from the structure
great success [27], such as computer vision (CV), natural of DBNs, such as fault detection classification in industrial
language processing (NLP), and artificial intelligence (AI). environments, threat identification in security alert systems,
Compared to traditional machine learning methods, deep and emotional feature extraction out of images [17].
learning has demonstrated powerful information extraction
and processing capabilities, but also requires massive com- B. AUTOENCODER
putation resources. The breakthroughs of deep learning have An autoencoder includes an input layer and an output layer
greatly expanded the edge computing applications in various that are connected by one or multiple hidden layers [32],
scenarios, improving performance, efficiency, and manage- as illustrated in Fig. 2(b). The shape of the input layer and
ment. In this section, we introduce some typical deep learning the output layer are the same. The AE can be divided into
models that are widely used for edge computing applications, two parts, i.e., an encoder and a decoder. The encoder learns
including restricted Boltzmann machine (RBM), autoencoder the representative characteristics of the input and transforms
(AE), deep neural network (DNN), convolutional neural net- it into other latent features (usually in a compressing way).
4 VOLUME 4, 2016
And the decoder receives the latent features of the encoder control the flow of information. In LSTM model, the forget
and aims to reconstruct the original form of the input data, gate is utilized to control the cell state and decide what to
minimizing the reconstruction error. Similarly, an AE can be keep in the memory. Through the learning process, stored
formed as a deep architecture by stacking multiple layers into computations in the memory cells are not distorted over time,
the hidden layer. There are several variants and extensions which particularly achieves better performance when data is
of AEs, such as sparse AE [33], denoising AE [34], and characterized in long dependency [42]. RNN and LSTM are
variational AE [35]. widely used in various sequential scenarios, such as language
processing [43] and activity recognition [44].
C. DEEP NEURAL NETWORKS
Compared to the traditional artificial neural network (ANN) F. DEEP REINFORCEMENT LEARNING
that has shallow structure, deep neural network (DNN) (or Deep reinforcement learning (DRL) [7] is a combination of
deep fully connected neural network) usually has a deeper deep learning (DL) and reinforcement learning (RL) [45]. It
layer structure for more complicated learning tasks [32]. A aims to build an agent that is able to learn the best action
DNN consists of an input layer, several hidden layers, and choices over a set of states through the interaction with the
an output layer, where the output of each layer is fed to the environment, so as to maximize the long-term accumulated
next layer with activation functions. At the last layer, the rewards. Different from traditional RL, DRL utilizes a deep
final output representing the model prediction is produced. neural network to represent the policy given its strong rep-
Optimization algorithms such as Stochastic Gradient Decent resentation ability to approximate the value function or the
(SGD) [36] and backpropagation [37] are mostly used in the direct strategy. DRL can be categorized into value-based
training process. DNNs are widely used in feature extraction, models, such as Deep Q-Learning (DQL), Double DQL [46]
classification and function approximation. and Duel DQL [47], and policy-gradient-based models, such
as deep deterministic policy gradient (DDPG) [48] and asyn-
D. CONVOLUTIONAL NEURAL NETWORKS chronous advantage actor-critic (A3C) [49]. The DRL has
Convolutional neural networks (CNNs) are designed to pro- been successfully applied in many fields, such as computer
cess data that comes in the form of multiple arrays, for exam- gaming [7], chess gaming [8] and rate adaptation [50].
ple, a color image composed of three 2D arrays containing
pixel intensities in the three color channels [27]. A CNN IV. EMPOWERING EDGE APPLICATIONS WITH DEEP
receives 2D data structures and extracts high-level features LEARNING
through convolutional layers as described in Fig. 2(c), which A. WHEN EDGE COMPUTING MEETS DEEP LEARNING
is the core of CNN architecture. By going through the 2D Recent years have witnessed the rapid development and the
data with a set of moving filters and the pooling functions, achieved great success of edge computing and deep learning
CNN extracts the spatial correlations between adjacent data in their respective fields. However, the massive amount of
by calculating the inner product of the input and the filter. Af- invaluable data generated and collected at the edge side calls
ter that, a pooling block is operated over the output to reduce for more powerful and intelligent processing capacities lo-
the spatial dimensions and generate a high-level abstraction. cally to fully unleash the underlying potentials of big data, so
Compared to traditional fully connected deep networks, CNN as to satisfy the ever-increasing demands of various applica-
can effectively decrease the parameter numbers of network tions. Fortunately, the recent breakthroughs in deep learning
and extract the spatial correlations in the raw data, mitigating shed a light on the edge application scenarios, providing
the risk of overfitting [38]. The above advantages make strong ability in information perception, data management,
CNN achieve significant results in many applications, such decision making, etc. The convergence of these two tech-
as object detection [39] and health monitoring [40]. nologies can further create new opportunities, empowering
the development of many emerging applications. In fact,
E. RECURRENT NEURAL NETWORKS edge computing has already been gradually integrated with
Different from CNNs that are good at abstracting spatial artificial intelligence (AI) to achieve edge intelligence. In this
features, recurrent neural networks (RNNs) are designed for rest of this section, we conduct a comprehensive overview
processing sequential or time-series data. The input to an of state-of-the-art research works on edge computing appli-
RNN includes both the current sample and the previously cations with deep learning and summarize them in several
observed samples. Specifically, each neuron of an RNN layer aspects, including smart multimedia, smart transportation,
not only receives the output of its previous layer but also smart city, and smart industry.
receives the stored state of from previous time steps, as
depicted in Fig. 2(e). With this special architecture, RNN is B. SMART MULTIMEDIA
able to remember previous information for integrated pro- The Internet video content has been explosively increasing
cessing with the current information. However, RNNs can in the past years. As estimated by Cisco, the global video
only look back for a few steps due to the gradient explosion traffic accounted for 75% of the Internet traffic in 2017
and long-term dependencies. To solve this problem, Long and is estimated to grow four-fold by 2022 [63]. Mean-
Short-Term Memory (LSTM) network [41] is proposed to while, people are having an increasingly higher demand for
VOLUME 4, 2016 5
Application Work Model Summary

Ren et al. [51] R-CNN An objection detection architecture and implementation that leverages Faster R-
CNN for objection detection.
Video Analytics Liu et al. [52] CNN A CNN-based visual food recognition algorithms with edge computing.
DeepDecision [53] CNN, Yolo [54] A distributed infrastructure that ties end and edges to improve Yolo with higher
frame rate and accuracy.
DeepCham [55] CNN Use edge and end user to collaboratively train CNN model for better object
recognition accuracy.
Nikouei et al. [56] L-CNN Develop L-CNN for resource-constrained edge devices with reduced filter num-
bers.
Grazia [57] RBM, Liner Classifier A multi-stage learning system integrating RMB and liner classifier for simulta-
Adaptive Streaming neous video transmission with guarantees for each user.
Wang et al. [58] DRL An edge computing-assisted framework that leverages DRL to intelligently
assign user to proper edge servers.
Li et al. [59] ARIMA, MLR,kNN Consider the video propagation and popularity evolution patterns with ARIMA,
MLP, and kNN.
Caching Zhang et al. [60] LSTM-C Propose a caching framework LSTM-C to better learn to content popularity
patterns both at long and short time scale.
Zhu et al. [61] DRL Leverage DRL to automatically learn an end-to-end caching policy.
Jiang et al. [62] MADRL Formulate the D2D caching problem as a multi-agent MAB problem and
developing a multi-agent DRL solution.
TABLE 1: Summary of deep-learning-enabled edge computing applications in the field of smart multimedia.
video content and video watching experience, calling for detection, with a well-designed RoI detection algorithm to
more intelligent video processing, caching, and delivery, etc. balance the detection accuracy and the data compression
Nowadays, deep learning is integrated with edge computing rate. Liu et al. [52] developed a CNN-based visual food
to provide both better video quality of experience (QoE) for recognition algorithms to achieve the best-in-class recog-
viewers and cost-effective functions to service providers. The nition accuracy, where edge computing was employed to
representative researches on smart multimedia applications overcome the system latency and low battery life of mobile
are summarized in Tab. 1. devices. In DeepDecision [53], a distributed infrastructure
was proposed to tie together computationally weak frontend
1) Video Analytics devices (assumed to be smartphones) with more powerful
Nowadays, video analytics [9] are becoming more and more back-end helpers (the edges) to allow deep learning to choose
widely used in different fields such as camera-based surveil- local or remote execution. This approach boosts the perfor-
lance [64] and augmented reality (AR). With the limited mance of CNN, in particular Yolo [54], to achieve higher
processing capabilities of cameras, traditional video analytics frame rate and accuracy. DeepCham [55] leveraged an edge
usually heavily rely on cloud computing for content process- master server coordinated with several participating users
ing, i.e., the video contents are first streamed to the backend to collaboratively train CNN model to achieve better object
cloud servers and the processed results are then delivered to recognition accuracy. LAVEA [66] built up a client-edge
the frontend devices. This processing mode however brings collaboration system and solved an optimization offloading
high latency and consumes much bandwidth, unable to sat- problem to minimize the response time. Nikouei et al. [56]
isfy those latency-sensitive applications, not to mention those developed a lightweight convolutional neural network (L-
realtime requirements such as object detection [51] and track- CNN), which leveraged the depthwise separable convolution
ing [65]. The emergence of edge computing pushes the video feature and tailored the CNN to be furnished in the resource-
analytics from the remote cloud to the local edge, allowing constrained edge devices with reduced filter numbers in each
the video content to be processed near the data source so layer.
as to enable quick or even realtime response. For example,
Amazon has released the world’s first deep-learning-enabled 2) Adaptive Streaming
video camera, where the locally executed deep learning func- Adaptive video streaming [78] is becoming a critical issue
tion enables realtime objection even without the involvement in today’s video delivery to provide the best quality of ex-
of the cloud. perience (QoE) over the Internet. The basic idea is to select
Pioneer researches have conducted efforts towards intel- proper video bitrate considering the network states, stability,
ligent video analytics with edge computing. Ren et al. [51] fairness, user’s preference of video quality, etc. Most existing
proposed an edge-computing-based object detection archi- adaptive streaming approaches [79]–[81] rely on the client-
tecture as well as a preliminary implementation to achieve based adaptation, which aims to adapt to the bandwidth vari-
distributed and efficient object detection via wireless com- ations based on several observed or predicted metrics such
munications for real-time surveillance applications. They as buffer size and bandwidth situation. The recent advances
adopted Faster R-CNN [39] for model training and object in deep learning, particularly deep reinforcement learning,
6 VOLUME 4, 2016

SqueezeDet [67] CNN A careful designed CNN for objection with reduced model size.
Chen et al. [68] CNN A CNN-based monocular 3D object detection method based on the fact that objects
Autonomous Driving should be on the ground-plane.
MV3D [69] Fusion VGG A multi-view 3D deep learning network that takes both LIDAR point and camera
images as a fusion input.
Bojarski et al. [70] CNN An end-to-end learning architecture that directly mapped the raw pixels from a single
front-facing camera to the steering commands.
Polson et al. [71] DNN Leveraging deep neural network to mine the short term characteristics of the traffic
situation for traffic prediction.
Traffic Analysis Lv et al. [72] SAE Leverage a stacked autoencoder (SAE) to learn the generic traffic features from the
and Prediction historical data.
Koesdwiady et al. [73] DBN Consider the impact of weather on traffic situation and incorporat the weather
information into a deep belief network.
DeepTransport [74] LSTM Consider the mobility analysis at the citywide scale with a LSTM model for
prediction.
Yao et al. [75] CNN,LSTM Revisit the spatiotemporal relationships in traffic pattern and propose a novel spatial-
temporal dynamic network (STDN) based on CNN and LSTM for prediction.
Wiering et al. [76] Tabular QL Leverage a tabular Q-learning model in an isolated intersection for signal control.
Traffic Signal Control Abdoos et al. [10] MAQL Propose a multi-agent Q-learning (MAQL) method that considers the queue length
for cooperative scheduling.
Chu et al. [77] MA2C Propose a novel multi-agent actor-critic (MA2C) approach to comprehensively
combine the traffic features for intelligent control.
TABLE 2: Summary of deep-learning-enabled edge computing applications in the field of smart transportation.
are leveraged at the client end to automatically learn the lutions such as Least Recently Used (LRU), Least Frequently
adaptation policy for better QoE [50], [82]. Used (LFU) and their variants [88], [89], or model-based
The emergence of edge computing provides an alternative solutions such as [90], [91] given the easy implementation.
for adaptive video streaming. Given the intelligence and These solutions however heavily rely on dedicated features
computation capability, edge nodes can serve as a cache and are not adaptive enough to the changing characteristics.
server or transcoding server to provide edge-based (or edge- Deep learning brings new opportunities towards intelligent
cloud-based) adaptive streaming [83], [84]. This scheme can edge caching using advanced learning techniques to well
usually collect more states from other users and achieve capture the hidden features. Li et al. [59] considered the video
higher fairness, stability, and collaborative intelligence. For propagation as well as popularity evolution patterns and de-
example, De Grazia [57] developed a multi-stage learning veloped an integration of ARIMA, multiple linear regression
system to manage simultaneous video transmission which (MLR), and k-nearest neighbor regression (kNN) to predict
guarantees a minimum quality level for each user. In par- the social patterns to improve caching performance. Zhang et
ticular, they used an unsupervised Restricted Boltzmann al. [60] proposed an intelligent edge-assisted caching frame-
Machine (RBM) [28] to capture the latent features of the work LSTM-C based on LSTM to better learn to content
input data and a supervised linear classifier to estimate the popularity patterns both at long and short time scale. Zhu et
characteristics of unknown videos. Wang et al. [58] designed al. [61] proposed to leverage DRL to automatically learn an
an edge computing-assisted framework that leverages DRL end-to-end caching policy, where the user requests, network
to intelligently assign users to proper edge servers to achieve constraints, and external information are all embedded in the
proper video streaming services. learning environment. Besides individual caching decisions,
collaborative caching is also explored in recent years to
3) Caching achieve collective intelligence. For example, Jiang et al. [62]
formulates the D2D caching problem as a multi-agent multi-
Video content caching [85] is another important applica-
armed bandit (MAB) problem and developed a multi-agent
tion that has attracted continuous research efforts for years
DRL (MADRL) solution to learn a coordinated caching
given its great benefits in improving multimedia services.
scheme among multiple agents.
In traditional content delivery network (CDN) architecture,
video contents are usually placed or cached at remote servers,
where the high latency and limited bandwidth between the C. SMART TRANSPORTATION
end viewers and the remote servers can cause viewing delay Vehicle is envisioned as the next intelligent information
and congestion, seriously undermining the viewers’ QoE. carrier after smartphone. The coming era of 5G and mobile
The emerging edge caching [86], [87] is able to alleviate this edge computing (MEC) has enabled vehicle information to
problem by pushing the content close to the end users so as be readily accessible anytime and anywhere with low latency,
to reduce the access latency and reduce the network pressure. forming an Internet of Vehicle (IoV) [92]. Integrated with
Traditional content providers may simply use rule-based so- the latest advances in deep learning, IoV will enable more
VOLUME 4, 2016 7
intelligent transportation management, such as autonomous ered the mobility analysis at a larger scale, i.e., the citywide
driving [24], traffic prediction, traffic signal control, as sum- scale. LSTM model is used for future movement prediction.
marized in Tab. 2. In [75], the authors revisited the spatiotemporal relationships
in traffic patterns and proposed a novel spatial-temporal
1) Autonomous Driving dynamic network (STDN) based on CNN and LSTM, which
Intelligent sensing and perception are of the most critical outperforms the existing prediction methods.
issues in autonomous driving [24]. The vehicles first collect
the information from various carried sensors such as cameras 3) Traffic Signal Control
and radars, and then conduct an intelligent perception and de- With the above traffic analysis and prediction, the combina-
cision. Purely using vehicle-based and cloud-based solutions tion of edge computing and deep learning actually can do
may not well satisfy the requirement of high computation more things towards intelligent transportation management.
capacity, realtime feedback, enough redundancy, and security Among them, intelligent traffic signal control [76] is one
for autonomous driving. Edge computing however provides of the most representative applications and has also been
a promising solution with powerful computation and low- explored by researchers for years. A good control policy is
latency communication [105]. With the benefits of V2X able to reduce the average waiting time, traffic congestion,
communications [106], part of the learning-based perception and traffic accident. The early traffic signal control methods
can be offloaded to the edge server for processing. usually rely on fuzzy logic [111] or genetic algorithm [112].
Many existing works have conducted pioneer efforts to- The key challenge however lies in how to achieve collabora-
wards autonomous driving. SqueezeDet [67] proposed a care- tive and intelligent control among multiple or even citywide
fully designed CNN-based learning pipeline that not only traffic lights for large scale traffic scheduling. Towards this
achieves high object detection accuracy but also reduces the goal, reinforcement learning (RL) and multi-agent RL turns
model size for energy efficiency. To better understand the out to be a promising solution where each agent (can be
captured object, Chen et al. [68] proposed a monocular 3D implemented as an edge) will make control policy for a traffic
object detection method using state-of-the-art CNN method light considering not only its local traffic situation but also
based on the fact that objects should be on the ground-plane. other agents’ traffic situations. In [76], tabular Q-learning
To further improve the accuracy and robustness, MV3D [69] was first applied in an isolated intersection for signal control.
developed a multi-view 3D deep learning network that takes To improve the collaboration among traffic lights, Abdoos et
both LIDAR point and camera images as a fusion input to al. [10] proposed a multi-agent Q-learning (MAQL) method
predict 3D boundaries. Besides the road feature extraction, that considered the queue length for cooperative scheduling.
researchers also dived deeply into the driving control based The latest work [77] further integrated the state-of-the-art
on the sensing information. Bojarski et al. [70] proposed actor-critic (A2C) RL algorithm and the multi-agent learning
an end-to-end learning architecture without detecting road as a multi-agent actor-critic (MA2C) approach to comprehen-
features. It directly mapped the raw pixels from a single sively combine the traffic features for intelligent control.
front-facing camera to the steering commands.
D. SMART CITY
2) Traffic Analysis and Prediction Smart city [113] is another important application scenario for
Understanding the mobility patterns of the vehicles and deep-learning-enabled edge computing. The geo-distributed
people is a critical problem for urban traffic management, big data [114] in a city naturally requires a distributed com-
city planning, and service provisioning. Given the distributed puting paradigm for local processing and management. The
features of mobile edge servers, edge computing is naturally integration of edge computing and deep learning enables the
ideal for vehicle traffic analysis and prediction [107]. Tra- deep penetration of computing intelligence into every corner
ditional approaches mostly used time-series analysis [108] or of a city, forming a smart city that can provide more efficient,
probabilistic graph analysis [109], which may not sufficiently economic, energy-saving, and convenient services [115],
capture the hidden spatiotemporal relationships therein. As [116]. We next discuss the combinational advantages in the
a powerful learning tool, deep learning stands out as an main components of the smart city, including the smart home,
effective method in this direction. Liu et al. [110] further smart building, and smart grid, as in Tab. 3.
pointed out the potential of applying different deep learning
approaches in urban traffic prediction. Polson et al. [71] 1) Smart Home
leveraged deep neural network to mine the short term charac- Smart IoT has been widely explored in smart home scenarios
teristics of the traffic situation of a road segment to predict the to provide not only convenient control but also intelligent
near future traffic pattern. Lv et al. [72] leveraged a stacked sensing [117]. Considering the privacy issue in the home
autoencoder (SAE) to learn the generic traffic features from scenario, edge computing is a good choice to provide local
the historical data. Koesdwiady et al. [73] further considered computation and processing, especially for the computation-
the impact of weather on traffic situations and incorporated intensive deep-learning-based applications. Dhakal et al. [93]
the weather information into a deep belief network for inte- developed an automated home/business monitoring system
grated learning. In DeepTransport [74], the authors consid- which resides on Network Function Virtualization (NFV)
8 VOLUME 4, 2016

Dhakal et al. [93] kNN,DNN Develop an automated home/business monitoring system on NFV edge servers
performing online learning on streaming data from home.
Smart Home SignFi [94] CNN Exploit the CSI signals of WiFi and using CNN to identify 276 sign language
gestures.
Wang et al. [95] CNN,LSTM Leverage a combined CNN and LSTM to recognize different gestures and
activities.
Mohammadi et al. [96] DRL Propose a novel semisupervised DRL-based method for indoor localization.
Zheng et al. [97], [98] Multi-task Learning Leverage multi-task learning to predict the performance of a chiller and strike
Smart Building balance between electricity consumption and real-world deployment.
Yuce et al. [99] NN Propose a neural network based model to perform regression analysis of energy
consumption within a building.
Thokala [100] SVM,RNN Consider the heterogeneity in the electrical load and propose to use both SVM
and partial RNN to forecast future load.
He et al. [101] DBN,RBM A deep learning based mechanism that integrates DBN and RBM to detect the
attack behavior of false data injection in realtime.
Smart Grid Yan et al. [102] RL A reinforcement learning-based approach to identify critical attack sequences
with consideration of physical system behaviors.
Shi et al. [103] RNN A novel pooling-based RNN network to forecast the household load addressing
the over-fitting issue.
Wan et al. [104] DRL A model-free DRL-based model to automatically determine the charging policy.
TABLE 3: Summary of deep-learning-enabled edge computing applications in the field of smart city.
edge servers performing online learning on streaming data of energy consumption within a building. Thokala [100]
coming from homes and businesses in the neighborhood. further considered the heterogeneity in the electrical load and
Leveraging the latest advances of deep learning, the ubiqui- proposed to use both SVM and partial RNN to forecast future
tous wireless signals can also be used for smart interaction load.
between human and devices. For example, SignFi [94] ex-
ploited the CSI signals of WiFi and was able to identify 276 3) Smart Grid
sign language gestures including the head, arm, hand, and The smart grid is an electricity distribution network with
finger with CNN for classification. Wang et al. [95] analyzed smart meters deployed at various locations to measure the
the impact patterns of moving humans on the WiFi signals realtime status information [127]. Smart grid is also an
and leveraged a combined CNN and LSTM to recognize important use case for edge computing or fog computing.
different gestures and activities. Such method can be used Edge collectors at the edge ingest the data generated by
for remote control of home devices such as lights and tele- grid sensors and devices, where some data for protection
visions [118]. Mohammadi et al. [96] explored more pos- and control loops even require real-time processing (from
sibilities of deep learning approaches and proposed a novel milliseconds to sub-seconds) [3]. Deep learning together with
semisupervised DRL based method for indoor localization. the edge computing empowers the grid with more intelli-
Such edge-intelligence-enabled solution can be widely used gent protection, control, and management. He et al. [101]
for smart home, including intrusion detection, gesture-based proposed a deep-learning-based mechanism that integrated
interaction, fall detection, etc. deep belief network (DBN) and RBM to detect the attack
behavior of false data injection in realtime. Considering the
2) Smart Building sequential behavior of attacks for smart grid, Yan et al. [102]
Achieving intelligent monitoring, sensing, and control in the further proposed a reinforcement learning-based approach
building environment requires more comprehensive percep- to identify critical attack sequences with consideration of
tion and processing compared to the home environment given physical system behaviors. Shi et al. [103] proposed a novel
the complex architecture. In this context, edge computing pooling-based RNN network to forecast the household load
plays an important role in data processing, orchestration, and addressing the over-fitting issue. Pricing is another important
privacy preserving. Recently, many efforts [119]–[121] have issue towards smart grid, which greatly affects customers’
been made towards smart building to reduce energy con- using behaviors in many aspects, e.g., the economy-driven
sumption, improve the building security, enhance the sensing electric vehicle charging [128]. For instance, Wan et al. [104]
capacity of buildings, etc. Zheng et al. [97], [98] focused jointly considered the electricity price and battery energy
on the chiller sequencing problem to reduce the electricity of electric vehicles and proposed a model-free DRL based
consumption in buildings. They leveraged multi-task learning model to automatically determine the charging policy.
to predict the performance of a chiller and further strike a
good balance between the electricity consumption and ease E. SMART INDUSTRY
of use for real-world deployment. Yuce et al. [99] proposed a In the coming era of industry 4.0, we are experiencing a rev-
neural-network-based model to perform regression analysis olution of smart industry, which has two main principles, i.e.,
VOLUME 4, 2016 9

Weimer et al. [122] CNN A novel CNN-based architecture for fast and reliable industrial inspection.
Smart Manufacturing Li et al. [123] CNN An edge-computing-based model that offloads computation burden to the fog
nodes and a CNN-based model with an early-exit design for accurate inspection.
Zhao et al. [124] CNN, bi-LSTM A system that combined CNN and bi-directional LSTM for machine health
monitoring.
Wu et al. [125] vanilla LSTM A remaining life prediction for engineered system using vanilla LSTM neural
Smart Industrial Analysis networks.
Wang et al. [126] DNN Proposing a DNN-based architecture to accurately predict the remaining energy
and remaining lifetime of batteries, which further enables an informed power
configuration among base stations.
TABLE 4: Summary of deep-learning-enabled edge computing applications in the field of smart industry.
production automation and smart data analysis [123]. The V. RESEARCH CHALLENGES AND DIRECTIONS
former is one of our main objectives that could greatly liber- Though the convergence of edge computing and deep learn-
ate the productivity and the latter is one of the most effective ing has revealed great potentials and prompted the fast
methods towards our objectives. The recent advances of edge development of many applications, there still exist various
computing migrate the massive computation from the remote problems in achieving stable, robust, and practical usage,
cloud to the local edge, enabling more low-latency and secure which calls for continuous efforts in this field from many
manufacturing [129]. And deep learning further empowers perspectives. We next discuss some key research challenges
more effective local analysis and prediction [130] at the edge and promising directions.
node of industry instead of the cloud. We summarize them in
the next two parts as illustrated in Tab. 4. A. MODEL TRAINING
The performance of deep-learning-enabled edge applications
1) Smart Manufacturing highly relies on how the learning models are performed
Smart manufacturing is the key component of the smart in the edge computing architecture, where the model train-
industry, which highly relies on the intelligent processing of ing is an important process. It is well known that model
deep learning and quick response of edge computing. The training is often computation-intensive, consuming massive
combination of deep learning and edge computing has been CPU and GPU resources, especially for those deep models.
applied in many aspects of industry manufacturing, such Edge servers are usually challenging, or at least not cost-
as manufacture inspection and fault assessment. Weimer et efficient, to solely take the model training tasks. Besides,
al. [122] developed a novel CNN-based architecture for fast in many applications, the data is distributed collected from
and reliable industrial inspection, which can automatically multiple edge servers and it is difficult for a single edge
generate meaningful features for a specific inspection task server to obtain the whole information for model training.
from a huge amount of raw data with minimal human in- Sharing the raw data among the edge nodes is not a good
teraction. Li et al. [123] proposed an edge-computing-based solution since it will consume massive communication re-
model that is able to offload the computation burden to sources. Towards this direction, distributed learning [131]
the fog nodes to deal with extremely large data. A CNN- and federated learning [132] are two promising models to
based model together with an early-exit design is used in address this problem. The early idea of distributed learning is
this model, which largely improved the inspection accuracy to design a decentralized Stochastic Gradient Descent (SGD)
and robustness. Zhao et al. [124] further combined CNN algorithm in the edge computing environment. The key chal-
with bi-directional LSTM to propose a novel machine health lenge exists in reducing the communication cost for gradient
monitoring system. updates while preserving the training accuracy. Recent efforts
have been made towards this direction, e.g., delivering the
2) Smart Industrial Analysis important gradient first [133]. Federated learning is another
Besides the manufacture inspection and monitoring, the ap- promising method emerging in recent years to train deep
plication of edge computing and deep learning also enables neural networks as it leaves the raw data on clients and only
much intelligent industrial analysis. Wu et al. [125] focused aggregates the intermediate updates from each client. In the
on the remaining useful life estimation of the engineered edge application scenario, it can also reduce the communica-
system and proposed to use vanilla LSTM neural networks to tion cost and improve resource utilization [134].
get good remaining lifetime prediction accuracy. Wang et al.
[126] focused on the remaining lifetime analysis of backup B. MODEL INFERENCE
batteries in the wireless base stations. They proposed to use As many outstanding learning models are getting bigger
DNN-based architecture to accurately predict the remaining and deeper, traditional large-scale learning models are often
energy and remaining lifetime of batteries, which further en- deployed in a centralized cloud and receive the raw input
ables an informed power configuration among base stations. data from the distributed end devices, which can cause high
10 VOLUME 4, 2016
delay. Edge servers provide alternative solutions for model such as Amazon’s Greengrass2 and Microsoft’s Azure IoT
deployment, where the edge and cloud can work collabora- Edge3 . For the perspective of programming abstraction, there
tively to handle the massive amounts of learning takes. A are also some frameworks specially designed for edge sce-
promising direction is model partition for deep neural net- nario, such as MXNet [140], Tensorflow Lite4 and CoreML5 .
works, where the end, edge, and cloud will execute part of the Though with these existing systems and frameworks, there
learning models, respectively. For example, Kang et al. [135] still need a lot of efforts to integrate them to achieve a
developed a lightweight scheduler to automatically partition more practical and high-performance system for general edge
DNN computation between mobile devices and data centers learning applications.
at the granularity of neural network layers. And Huang et
al. [136] further explored the partitioning problem in the edge VI. CONCLUSION
computing scenario, and developed a partitioning strategy In this article, we mainly investigated how the recent ad-
among the device, edge, and cloud, which aimed to reduce vances of deep learning can be leveraged to improve the
the execution delay. Another promising direction is the early novel edge computing applications. We first introduced the
exit of inference (EEoI) [137] for deep neural networks. basic concepts and paradigms of edge computing, highlight-
Since passing through the whole deep networks is both time- ing its key advantages. We then present some representative
and energy-consuming for edge servers, EEoI allows the deep learning models that can be used in edge computing,
inference to exit early if verified by some predefined models. such as autoencoder, CNN, RNN, DRL, etc. A comprehen-
sive survey on the latest deep learning empowered edge
C. APPLICATION ENHANCEMENT computing applications is next conducted from four domains,
including smart multimedia, smart transportation, smart city,
The integration of deep learning and edge computing has and smart industry. Finally, we discussed the key challenges
achieved remarkable improvement for many application sce- and future research directions on improving the intelligent
narios, yet there are still some critical applications desir- edge computing applications. We hope this survey is able
ing for real breakthroughs. Real-time VR gaming and au- to elicit more discussion and inspiration on the convergence
tonomous driving are two most representative applications. of edge computing and deep learning, so as to facilitate
Both these two applications require ultra low-delay interac- the development of deep-learning-enabled edge computing
tions and powerful computations. The emerging 5G commu- applications.
nication technology together with edge learning brings new
possibilities towards a feasible solution, where the main com-
REFERENCES
putation such as video rendering as well as video analytics
[1] “Fog computing and the internet of things: Extend the cloud to where the
can be conducted at local edges and the processing results can things are, white paper,” https://www.cisco.com/c/dam/en_us/solutions/
be delivered to the end in near real time, e.g., millisecond- trends/iot/docs/computing-overview.pdf.
level interaction. Despite of the preliminary foundation of [2] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, “Edge computing: Vision and
challenges,” IEEE Internet of Things Journal, vol. 3, no. 5, pp. 637–646,
feasibility, there is still a long way to go before the practical 2016.
application. [3] F. Bonomi, R. Milito, J. Zhu, and S. Addepalli, “Fog computing and its
role in the internet of things,” in Proceedings of ACM MCC Workshop.
ACM, 2012, pp. 13–16.
D. HARDWARE AND SOFTWARE OPTIMIZATION [4] M. Satyanarayanan, V. Bahl, R. Caceres, and N. Davies, “The case for
vm-based cloudlets in mobile computing,” IEEE pervasive Computing,
In addition to the model and application enhancement, the 2009.
system-level optimization for deep learning and edge com- [5] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “Imagenet classification
with deep convolutional neural networks,” in Proceedings of NeurIPS,
puting is also a challenging yet promising direction. Most 2012, pp. 1097–1105.
of the existing hardware architecture, software platform, [6] T. Young, D. Hazarika, S. Poria, and E. Cambria, “Recent trends in
and programming abstraction are particularly designed for deep learning based natural language processing,” IEEE Computational
Intelligence Magazine, vol. 13, no. 3, pp. 55–75, 2018.
the cloud-based computing paradigm. Yet the edge learning
[7] V. Mnih, K. Kavukcuoglu et al., “Human-level control through deep
emphasizes some different aspects compared to cloud com- reinforcement learning,” Nature, vol. 518, no. 7540, p. 529, 2015.
puting, such as energy-efficiency, light-weight architecture, [8] D. Silver, A. Huang et al., “Mastering the game of go with deep neural
and edge-oriented computation framework. For example, networks and tree search,” Nature, vol. 529, no. 7587, p. 484, 2016.
[9] G. Ananthanarayanan, P. Bahl, P. Bodík, K. Chintalapudi, M. Philipose,
from the perspective of the hardware architecture optimiza- L. Ravindranath, and S. Sinha, “Real-time video analytics: The killer app
tion, Du et al. [138] studied Cortex-M micro-controllers for edge computing,” IEEE Computer, vol. 50, no. 10, pp. 58–67, 2017.
and proposed a streaming hardware accelerator to better [10] M. Abdoos, N. Mozayani, and A. L. Bazzan, “Traffic light control in non-
stationary environments based on multi agent q-learning,” in Proceedings
CNN in edge devices. Besides, FPGA-based edge comput- of IEEE ITSC, 2011, pp. 1580–1585.
ing platforms are also developed to support deep learning
computation offloading from mobile devices to the edge 2 https://aws.amazon.com/cn/greengrass/
FPGA platform [139]. For the software perspective, many 3 https://azure.microsoft.com/en-ca/services/iot-edge/
incorporations have proposed their own software platforms 4 https://www.tensorflow.org/lite
or services to support edge-level learning and computing, 5 https://developer.apple.com/documentation/coreml?language=objc
VOLUME 4, 2016 11
[11] Y. Han, X. Wang, V. Leung, D. Niyato, X. Yan, and X. Chen, “Conver- [37] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning represen-
gence of edge computing and deep learning: A comprehensive survey,” tations by back-propagating errors,” Nature, vol. 323, no. 6088, pp. 533–
arXiv preprint arXiv:1907.08349, 2019. 536, 1986.
[12] J. Chen and X. Ran, “Deep learning with edge computing: A review,” [38] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolu-
Proceedings of the IEEE, vol. 107, no. 8, pp. 1655–1674, 2019. tional networks,” in Proceedings of ECCV. Springer, 2014, pp. 818–
[13] Z. Zhou, X. Chen, E. Li, L. Zeng, K. Luo, and J. Zhang, “Edge intelli- 833.
gence: Paving the last mile of artificial intelligence with edge computing,” [39] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-cnn: Towards real-
arXiv preprint arXiv:1905.10083, 2019. time object detection with region proposal networks,” in Proceedings of
[14] S. Wang, X. Zhang, Y. Zhang, L. Wang, J. Yang, and W. Wang, “A NeurIPS, 2015, pp. 91–99.
survey on mobile edge networks: Convergence of computing, caching [40] N. Y. Hammerla, S. Halloran, and T. Plötz, “Deep, convolutional, and
and communications,” IEEE Access, vol. 5, pp. 6757–6779, 2017. recurrent models for human activity recognition using wearables,” arXiv
[15] N. Abbas, Y. Zhang, A. Taherkordi, and T. Skeie, “Mobile edge com- preprint arXiv:1604.08880, 2016.
puting: A survey,” IEEE Internet of Things Journal, vol. 5, no. 1, pp. [41] S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural
450–465, 2017. computation, vol. 9, no. 8, pp. 1735–1780, 1997.
[16] Y. Mao, C. You, J. Zhang, K. Huang, and K. B. Letaief, “A survey on [42] J. Chung, C. Gulcehre, K. Cho, and Y. Bengio, “Empirical evaluation of
mobile edge computing: The communication perspective,” IEEE Com- gated recurrent neural networks on sequence modeling,” arXiv preprint
munications Surveys & Tutorials, vol. 19, no. 4, pp. 2322–2358, 2017. arXiv:1412.3555, 2014.
[17] M. Mohammadi, A. Al-Fuqaha, S. Sorour, and M. Guizani, “Deep [43] Q. Chen, X. Zhu, Z. Ling, S. Wei, H. Jiang, and D. Inkpen, “Enhanced
learning for iot big data and streaming analytics: A survey,” IEEE lstm for natural language inference,” arXiv preprint arXiv:1609.06038,
Communications Surveys & Tutorials, vol. 20, no. 4, pp. 2923–2960, 2016.
2018. [44] Y. Guan and T. Plötz, “Ensembles of deep lstm learners for activity
[18] X. Ma, T. Yao, M. Hu, Y. Dong, W. Liu, F. Wang, and J. Liu, “A survey recognition using wearables,” Proceedings of ACM IMWUT, vol. 1,
on deep learning empowered iot applications,” IEEE Access, vol. 7, pp. no. 2, p. 11, 2017.
181 721–181 732, 2019. [45] R. S. Sutton, A. G. Barto et al., Introduction to reinforcement learning.
[19] T. Verbelen, P. Simoens, F. De Turck, and B. Dhoedt, “Cloudlets: Bring- MIT press Cambridge, 1998, vol. 2, no. 4.
ing the cloud to the mobile user,” in Proceedings of ACM MCS, 2012, [46] H. Van Hasselt, A. Guez, and D. Silver, “Deep reinforcement learning
pp. 29–36. with double q-learning,” in Proceedings of AAAI, 2016.
[20] K. Bilal, O. Khalid, A. Erbad, and S. U. Khan, “Potentials, trends, and [47] Z. Wang, T. Schaul, M. Hessel, H. Van Hasselt, M. Lanctot, and
prospects in edge technologies: Fog, cloudlet, mobile edge, and micro N. De Freitas, “Dueling network architectures for deep reinforcement
data centers,” Computer Networks, vol. 130, pp. 94–120, 2018. learning,” arXiv preprint arXiv:1511.06581, 2015.
[21] M. Aazam and E.-N. Huh, “Fog computing micro datacenter based
[48] T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y. Tassa, D. Silver,
dynamic resource estimation and pricing model for iot,” in Proceedings
and D. Wierstra, “Continuous control with deep reinforcement learning,”
of IEEE AINA, 2015, pp. 687–694.
arXiv preprint arXiv:1509.02971, 2015.
[22] F. Bonomi, R. Milito, P. Natarajan, and J. Zhu, “Fog computing: A
[49] V. Mnih, A. P. Badia et al., “Asynchronous methods for deep reinforce-
platform for internet of things and analytics,” in Big data and internet
ment learning,” in Proceedings of ICML, 2016, pp. 1928–1937.
of things: A roadmap for smart environments. Springer, 2014, pp. 169–
[50] H. Mao, R. Netravali, and M. Alizadeh, “Neural adaptive video streaming
186.
with pensieve,” in Proceedings of ACM SIGCOMM, 2017, pp. 197–210.
[23] Y. C. Hu, M. Patel, D. Sabella, N. Sprecher, and V. Young, “Mobile edge
computing - a key technology towards 5g,” ETSI white paper, vol. 11, [51] J. Ren, Y. Guo, D. Zhang, Q. Liu, and Y. Zhang, “Distributed and efficient
no. 11, pp. 1–16, 2015. object detection in edge computing: Challenges and solutions,” IEEE
Network, vol. 32, no. 6, pp. 137–143, 2018.
[24] S. Liu, L. Liu, J. Tang, B. Yu, Y. Wang, and W. Shi, “Edge computing for
autonomous driving: Opportunities and challenges,” Proceedings of the [52] C. Liu, Y. Cao, Y. Luo, G. Chen, V. Vokkarane, M. Yunsheng, S. Chen,
IEEE, vol. 107, no. 8, pp. 1697–1716, 2019. and P. Hou, “A new deep learning-based food recognition system for
[25] C. Perera, A. Zaslavsky, P. Christen, and D. Georgakopoulos, “Context dietary assessment on an edge computing service infrastructure,” IEEE
aware computing for the internet of things: A survey,” IEEE communica- Transactions on Services Computing, vol. 11, no. 2, pp. 249–261, 2017.
tions surveys & tutorials, vol. 16, no. 1, pp. 414–454, 2013. [53] X. Ran, H. Chen, X. Zhu, Z. Liu, and J. Chen, “Deepdecision: A mobile
[26] C. Wang, G. Liu, H. Huang, W. Feng, K. Peng, and L. Wang, “Miasec: deep learning framework for edge video analytics,” in Proceeding of
Enabling data indistinguishability against membership inference attacks IEEE INFOCOM, 2018, pp. 1421–1429.
in mlaas,” IEEE Transactions on Sustainable Computing, 2019. [54] J. Redmon and A. Farhadi, “Yolo9000: better, faster, stronger,” in Pro-
[27] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, ceedings of IEEE CVPR, 2017, pp. 7263–7271.
no. 7553, p. 436, 2015. [55] D. Li, T. Salonidis, N. V. Desai, and M. C. Chuah, “Deepcham: Collab-
[28] A. Fischer and C. Igel, “An introduction to restricted boltzmann ma- orative edge-mediated adaptive deep learning for mobile object recogni-
chines,” in Iberoamerican Congress on Pattern Recognition. Springer, tion,” in Proceedings of IEEE/ACM SEC, 2016, pp. 64–76.
2012, pp. 14–36. [56] S. Y. Nikouei, Y. Chen, S. Song, R. Xu, B.-Y. Choi, and T. Faughnan,
[29] R. Salakhutdinov, A. Mnih, and G. Hinton, “Restricted boltzmann ma- “Smart surveillance as an edge network service: From harr-cascade, svm
chines for collaborative filtering,” in Proceedings of ACM ICML, 2007, to a lightweight cnn,” in Proceedings of IEEE CIC, 2018, pp. 256–265.
pp. 791–798. [57] M. D. F. De Grazia, D. Zucchetto, A. Testolin, A. Zanella, M. Zorzi, and
[30] U. Fiore, F. Palmieri, A. Castiglione, and A. De Santis, “Network M. Zorzi, “Qoe multi-stage machine learning for dynamic video stream-
anomaly detection with the restricted boltzmann machine,” Neurocom- ing,” IEEE Transactions on Cognitive Communications and Networking,
puting, vol. 122, pp. 13–23, 2013. vol. 4, no. 1, pp. 146–161, 2017.
[31] G. E. Hinton, S. Osindero, and Y.-W. Teh, “A fast learning algorithm for [58] F. Wang, C. Zhang, J. Liu, Y. Zhu, H. Pang, L. Sun et al., “Intelligent
deep belief nets,” Neural computation, vol. 18, no. 7, pp. 1527–1554, edge-assisted crowdcast with deep reinforcement learning for personal-
2006. ized qoe,” in Proceedings of IEEE INFOCOM, 2019, pp. 910–918.
[32] I. Goodfellow, Y. Bengio, and A. Courville, Deep Learning. MIT Press, [59] H. Li, X. Ma, F. Wang, J. Liu, and K. Xu, “On popularity prediction of
2016, http://www.deeplearningbook.org. videos shared in online social networks,” in Proceedings of ACM CIKM,
[33] H. Lee, A. Battle, R. Raina, and A. Y. Ng, “Efficient sparse coding 2013, pp. 169–178.
algorithms,” in Proceedings of NeurIPS, 2007, pp. 801–808. [60] C. Zhang, H. Pang, J. Liu, S. Tang, R. Zhang, D. Wang, and L. Sun,
[34] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol, “Extracting “Toward edge-assisted video content intelligent caching with long short-
and composing robust features with denoising autoencoders,” in Proceed- term memory learning,” IEEE Access, vol. 7, pp. 152 832–152 846, 2019.
ings of ACM ICML, 2008, pp. 1096–1103. [61] H. Zhu, Y. Cao, W. Wang, T. Jiang, and S. Jin, “Deep reinforcement
[35] C. Doersch, “Tutorial on variational autoencoders,” arXiv preprint learning for mobile edge caching: Review, new features, and open issues,”
arXiv:1606.05908, 2016. IEEE Network, vol. 32, no. 6, pp. 50–57, 2018.
[36] L. Bottou, “Large-scale machine learning with stochastic gradient de- [62] W. Jiang, G. Feng, S. Qin, T. S. P. Yum, and G. Cao, “Multi-agent
scent,” in Proceedings of COMPSTAT. Springer, 2010, pp. 177–186. reinforcement learning for efficient content caching in mobile d2d net-
12 VOLUME 4, 2016
works,” IEEE Transactions on Wireless Communications, vol. 18, no. 3, [84] D. Wang, Y. Peng, X. Ma, W. Ding, H. Jiang, F. Chen, and J. Liu, “Adap-
pp. 1610–1622, 2019. tive wireless video streaming based on edge computing: Opportunities
[63] “Cisco visual networking index: Forecast and trends, and approaches,” IEEE Transactions on services Computing, vol. 12,
2017 to 2022 white paper,” https://www.cisco.com/c/en/us/ no. 5, pp. 685–697, 2018.
solutions/collateral/service-provider/visual-networking-index-vni/ [85] X. Wang, M. Chen, T. Taleb, A. Ksentini, and V. C. Leung, “Cache in the
white-paper-c11-741490.html. air: Exploiting content caching and delivery techniques for 5g systems,”
[64] P. Natarajan, P. K. Atrey, and M. Kankanhalli, “Multi-camera coordina- IEEE Communications Magazine, vol. 52, no. 2, pp. 131–139, 2014.
tion and control in surveillance systems: A survey,” ACM Transactions [86] E. Bastug, M. Bennis, and M. Debbah, “Living on the edge: The role
on Multimedia Computing, Communications, and Applications, vol. 11, of proactive caching in 5g wireless networks,” IEEE Communications
no. 4, p. 57, 2015. Magazine, vol. 52, no. 8, pp. 82–89, 2014.
[65] C.-C. Hung, G. Ananthanarayanan, P. Bodik, L. Golubchik, M. Yu, [87] E. Zeydan, E. Bastug, M. Bennis, M. A. Kader, I. A. Karatepe, A. S. Er,
P. Bahl, and M. Philipose, “Videoedge: Processing camera streams using and M. Debbah, “Big data caching for networking: Moving from cloud to
hierarchical clusters,” in Proceedings of IEEE/ACM SEC. IEEE, 2018, edge,” IEEE Communications Magazine, vol. 54, no. 9, pp. 36–42, 2016.
pp. 115–131. [88] S. Podlipnig and L. Böszörmenyi, “A survey of web cache replacement
[66] S. Yi, Z. Hao, Q. Zhang, Q. Zhang, W. Shi, and Q. Li, “Lavea: Latency- strategies,” ACM Computing Surveys, vol. 35, no. 4, pp. 374–398, 2003.
aware video analytics on edge computing platform,” in Proceedings of [89] Q. Huang, K. Birman, R. Van Renesse, W. Lloyd, S. Kumar, and H. C. Li,
ACM/IEEE SEC, 2017, p. 15. “An analysis of facebook photo caching,” in Proceedings of ACM SOSP,
[67] B. Wu, F. Iandola, P. H. Jin, and K. Keutzer, “Squeezedet: Unified, small, 2013, pp. 167–181.
low power fully convolutional neural networks for real-time object detec- [90] J. Hachem, N. Karamchandani, and S. Diggavi, “Content caching and
tion for autonomous driving,” in Proceedings of IEEE CVPR Workshops, delivery over heterogeneous wireless networks,” in Proceedings of IEEE
2017, pp. 129–137. INFOCOM, 2015, pp. 756–764.
[68] X. Chen, K. Kundu, Z. Zhang, H. Ma, S. Fidler, and R. Urtasun, [91] K. Shanmugam, N. Golrezaei, A. G. Dimakis, A. F. Molisch, and
“Monocular 3d object detection for autonomous driving,” in Proceedings G. Caire, “Femtocaching: Wireless content delivery through distributed
of IEEE CVPR, 2016, pp. 2147–2156. caching helpers,” IEEE Transactions on Information Theory, vol. 59,
[69] X. Chen, H. Ma, J. Wan, B. Li, and T. Xia, “Multi-view 3d object no. 12, pp. 8402–8413, 2013.
detection network for autonomous driving,” in Proceedings of the IEEE [92] F. Wang, F. Wang, X. Ma, and J. Liu, “Demystifying the crowd intelli-
CVPR, 2017, pp. 1907–1915. gence in last mile parcel delivery for smart cities,” IEEE Network, vol. 33,
[70] M. Bojarski, D. Del Testa, D. Dworakowski, B. Firner, B. Flepp, P. Goyal, no. 2, pp. 23–29, 2019.
L. D. Jackel, M. Monfort, U. Muller, J. Zhang et al., “End to end learning [93] A. Dhakal and K. Ramakrishnan, “Machine learning at the network
for self-driving cars,” arXiv preprint arXiv:1604.07316, 2016. edge for automated home intrusion monitoring,” in Proceedings of IEEE
ICNP, 2017, pp. 1–6.
[71] N. G. Polson and V. O. Sokolov, “Deep learning for short-term traffic flow
prediction,” Transportation Research Part C: Emerging Technologies, [94] Y. Ma, G. Zhou, S. Wang, H. Zhao, and W. Jung, “Signfi: Sign language
vol. 79, pp. 1–17, 2017. recognition using wifi,” Proceedings of ACM IMWUT, vol. 2, no. 1,
p. 23, 2018.
[72] Y. Lv, Y. Duan, W. Kang, Z. Li, and F.-Y. Wang, “Traffic flow prediction
[95] F. Wang, W. Gong, and J. Liu, “On spatial diversity in wifi-based human
with big data: a deep learning approach,” IEEE Transactions on Intelli-
activity recognition: A deep learning-based approach,” IEEE Internet of
gent Transportation Systems, vol. 16, no. 2, pp. 865–873, 2014.
Things Journal, vol. 6, no. 2, pp. 2035–2047, 2018.
[73] A. Koesdwiady, R. Soua, and F. Karray, “Improving traffic flow pre-
[96] M. Mohammadi, A. Al-Fuqaha, M. Guizani, and J.-S. Oh, “Semisu-
diction with weather information in connected cars: A deep learning
pervised deep reinforcement learning in support of iot and smart city
approach,” IEEE Transactions on Vehicular Technology, vol. 65, no. 12,
services,” IEEE Internet of Things Journal, vol. 5, no. 2, pp. 624–635,
pp. 9508–9517, 2016.
2017.
[74] X. Song, H. Kanasugi, and R. Shibasaki, “Deeptransport: Prediction and
[97] Z. Zheng, Q. Chen, C. Fan, N. Guan, A. Vishwanath, D. Wang, and F. Liu,
simulation of human mobility and transportation mode at a citywide
“Data driven chiller sequencing for reducing hvac electricity consump-
level.” in Proceedings of IJCAI, vol. 16, 2016, pp. 2618–2624.
tion in commercial buildings,” in Proceedings of ACM e-Energy, 2018,
[75] H. Yao, X. Tang, H. Wei, G. Zheng, and Z. Li, “Revisiting spatial- pp. 236–248.
temporal similarity: A deep learning framework for traffic prediction,” [98] Z. Zheng, Q. Chen, C. Fan, N. Guan et al., “An edge based data-driven
in Proceedings of AAAI, 2019. chiller sequencing framework for hvac electricity consumption reduction
[76] M. Wiering, J. v. Veenen, J. Vreeken, and A. Koopman, “Intelligent traffic in commercial buildings,” IEEE Transactions on Sustainable Computing,
light control,” 2004. 2019.
[77] T. Chu, J. Wang, L. Codecà, and Z. Li, “Multi-agent deep reinforcement [99] B. Yuce and Y. Rezgui, “An ann-ga semantic rule-based system to reduce
learning for large-scale traffic signal control,” IEEE Transactions on the gap between predicted and actual energy consumption in buildings,”
Intelligent Transportation Systems, 2019. IEEE Transactions on Automation Science and Engineering, vol. 14,
[78] A. Bentaleb, B. Taani, A. C. Begen, C. Timmerer, and R. Zimmermann, no. 3, pp. 1351–1363, 2015.
“A survey on bitrate adaptation schemes for streaming media over http,” [100] N. K. Thokala, A. Bapna, and M. G. Chandra, “A deployable electrical
IEEE Communications Surveys & Tutorials, vol. 21, no. 1, pp. 562–585, load forecasting solution for commercial buildings,” in Proceedings of
2018. IEEE ICIT, 2018, pp. 1101–1106.
[79] K. Spiteri, R. Urgaonkar, and R. K. Sitaraman, “Bola: Near-optimal [101] Y. He, G. J. Mendis, and J. Wei, “Real-time detection of false data injec-
bitrate adaptation for online videos,” in Proceedings of IEEE INFOCOM, tion attacks in smart grid: A deep learning-based intelligent mechanism,”
2016, pp. 1–9. IEEE Transactions on Smart Grid, vol. 8, no. 5, pp. 2505–2516, 2017.
[80] Y. Sun, X. Yin, J. Jiang, V. Sekar, F. Lin, N. Wang, T. Liu, and B. Sinopoli, [102] J. Yan, H. He, X. Zhong, and Y. Tang, “Q-learning-based vulnerability
“Cs2p: Improving video bitrate selection and adaptation with data-driven analysis of smart grid against sequential topology attacks,” IEEE Trans-
throughput prediction,” in Proceedings of ACM SIGCOMM, 2016, pp. actions on Information Forensics and Security, vol. 12, no. 1, pp. 200–
272–285. 210, 2016.
[81] T.-Y. Huang, R. Johari, N. McKeown, M. Trunnell, and M. Watson, “A [103] H. Shi, M. Xu, and R. Li, “Deep learning for household load forecasting
buffer-based approach to rate adaptation: Evidence from a large video - a novel pooling deep rnn,” IEEE Transactions on Smart Grid, vol. 9,
streaming service,” in ACM SIGCOMM Computer Communication Re- no. 5, pp. 5271–5280, 2017.
view, vol. 44, no. 4, 2014, pp. 187–198. [104] Z. Wan, H. Li, H. He, and D. Prokhorov, “Model-free real-time ev charg-
[82] T. Huang, R.-X. Zhang, C. Zhou, and L. Sun, “Qarc: Video quality aware ing scheduling based on deep reinforcement learning,” IEEE Transactions
rate control for real-time video streaming based on deep reinforcement on Smart Grid, 2018.
learning,” in Proceedings of ACM Multimedia, 2018, pp. 1208–1216. [105] Q. Yuan, H. Zhou, J. Li, Z. Liu, F. Yang, and X. S. Shen, “Toward efficient
[83] T. X. Tran, A. Hajisami, P. Pandey, and D. Pompili, “Collaborative content delivery for automated driving services: An edge computing
mobile edge computing in 5g networks: New paradigms, scenarios, and solution,” IEEE Network, vol. 32, no. 1, pp. 80–86, 2018.
challenges,” IEEE Communications Magazine, vol. 55, no. 4, pp. 54–61, [106] L. Hobert, A. Festag, I. Llatser, L. Altomare, F. Visintainer, and A. Ko-
2017. vacs, “Enhancements of v2x communication in support of cooperative
VOLUME 4, 2016 13
autonomous driving,” IEEE communications magazine, vol. 53, no. 12, [132] J. Konečnỳ, H. B. McMahan, F. X. Yu, P. Richtárik, A. T. Suresh, and
pp. 64–70, 2015. D. Bacon, “Federated learning: Strategies for improving communication
[107] I. Lana, J. Del Ser, M. Velez, and E. I. Vlahogianni, “Road traffic fore- efficiency,” arXiv preprint arXiv:1610.05492, 2016.
casting: recent advances and new challenges,” IEEE Intelligent Trans- [133] Y. Lin, S. Han, H. Mao, Y. Wang, and W. J. Dally, “Deep gradient
portation Systems Magazine, vol. 10, no. 2, pp. 93–109, 2018. compression: Reducing the communication bandwidth for distributed
[108] B. Ghosh, B. Basu, and M. O’Mahony, “Bayesian time-series model for training,” arXiv preprint arXiv:1712.01887, 2017.
short-term traffic flow forecasting,” Journal of transportation engineering, [134] W. Y. B. Lim, N. C. Luong, D. T. Hoang, Y. Jiao, Y.-C. Liang, Q. Yang,
vol. 133, no. 3, pp. 180–189, 2007. D. Niyato, and C. Miao, “Federated learning in mobile edge networks: A
[109] S. Sun and X. Xu, “Variational inference for infinite mixtures of gaussian comprehensive survey,” arXiv preprint arXiv:1909.11875, 2019.
processes with applications to traffic flow prediction,” IEEE Transactions [135] Y. Kang, J. Hauswald, C. Gao, A. Rovinski, T. Mudge, J. Mars, and
on Intelligent Transportation Systems, vol. 12, no. 2, pp. 466–475, 2010. L. Tang, “Neurosurgeon: Collaborative intelligence between the cloud
[110] Z. Liu, Z. Li, K. Wu, and M. Li, “Urban traffic prediction from mobility and mobile edge,” in Proceedings of ACM SIGARCH, vol. 45, no. 1,
data using deep learning,” IEEE Network, vol. 32, no. 4, pp. 40–46, 2018. 2017, pp. 615–629.
[111] B. P. Gokulan and D. Srinivasan, “Distributed geometric fuzzy multiagent [136] Y. Huang, F. Wang, F. Wang, and J. Liu, “Deepar: A hybrid device-edge-
urban traffic signal control,” IEEE Transactions on Intelligent Transporta- cloud execution framework for mobile deep learning applications,” in
tion Systems, vol. 11, no. 3, pp. 714–727, 2010. Proceedings of IEEE INFOCOM WKSHPS, 2019, pp. 892–897.
[112] H. Ceylan and M. G. Bell, “Traffic signal timing optimisation based on [137] S. Teerapittayanon, B. McDanel, and H.-T. Kung, “Branchynet: Fast
genetic algorithm approach, including drivers’ routing,” Transportation inference via early exiting from deep neural networks,” in Proceedings
Research Part B: Methodological, vol. 38, no. 4, pp. 329–342, 2004. of IEEE ICPR, 2016, pp. 2464–2469.
[113] F. Cicirelli, A. Guerrieri, G. Spezzano, and A. Vinci, “An edge-based plat- [138] L. Du, Y. Du, Y. Li, J. Su, Y.-C. Kuan, C.-C. Liu, and M.-C. F. Chang, “A
form for dynamic smart city applications,” Future Generation Computer reconfigurable streaming deep convolutional neural network accelerator
Systems, vol. 76, pp. 106–118, 2017. for internet of things,” IEEE Transactions on Circuits and Systems,
[114] X. He, K. Wang, H. Huang, and B. Liu, “Qoe-driven big data architecture vol. 65, no. 1, pp. 198–208, 2017.
for smart city,” IEEE Communications Magazine, vol. 56, no. 2, pp. 88– [139] S. Jiang, D. He, C. Yang, C. Xu, G. Luo, Y. Chen, Y. Liu, and J. Jiang,
93, 2018. “Accelerating mobile applications at the network edge with software-
[115] Y. Liu, C. Yang, L. Jiang, S. Xie, and Y. Zhang, “Intelligent edge com- programmable fpgas,” in Proceedings of IEEE INFOCOM, 2018, pp. 55–
puting for iot-based energy management in smart cities,” IEEE Network, 62.
vol. 33, no. 2, pp. 111–117, 2019. [140] T. Chen, M. Li, Y. Li, M. Lin, N. Wang, M. Wang, T. Xiao, B. Xu,
[116] M. Mohammadi and A. Al-Fuqaha, “Enabling cognitive smart cities C. Zhang, and Z. Zhang, “Mxnet: A flexible and efficient machine
using big data and machine learning: Approaches and challenges,” IEEE learning library for heterogeneous distributed systems,” arXiv preprint
Communications Magazine, vol. 56, no. 2, pp. 94–101, 2018. arXiv:1512.01274, 2015.
[117] H. Jiang, C. Cai, X. Ma, Y. Yang, and J. Liu, “Smart home based on wifi
sensing: A survey,” IEEE Access, vol. 6, pp. 13 317–13 325, 2018.
[118] Q. Pu, S. Gupta, S. Gollakota, and S. Patel, “Whole-home gesture
recognition using wireless signals,” in Proceedings of ACM MobiCom,
2013, pp. 27–38.
[119] W. Kleiminger, S. Santini, and F. Mattern, “Smart heating control with
occupancy prediction: how much can one save?” in Proceedings of ACM
Ubicomp, 2014, pp. 947–954.
[120] A. Vishwanath, V. Chandan, C. Mendoza, and C. Blake, “A data driven
pre-cooling framework for energy cost optimization in commercial build-
ings,” in Proceedings of ACM e-Energy, 2017, pp. 157–167.
[121] F.-J. Ferrández-Pastor, H. Mora, A. Jimeno-Morenilla, and B. Volckaert,
“Deployment of iot edge and fog computing technologies to develop
smart building services,” Sustainability, vol. 10, no. 11, p. 3832, 2018.
[122] D. Weimer, B. Scholz-Reiter, and M. Shpitalni, “Design of deep convo-
lutional neural network architectures for automated feature extraction in
industrial inspection,” CIRP Annals, vol. 65, no. 1, pp. 417–420, 2016.
[123] L. Li, K. Ota, and M. Dong, “Deep learning for smart industry: Efficient
manufacture inspection system with fog computing,” IEEE Transactions
on Industrial Informatics, vol. 14, no. 10, pp. 4665–4673, 2018.
[124] R. Zhao, R. Yan, J. Wang, and K. Mao, “Learning to monitor machine
health with convolutional bi-directional lstm networks,” Sensors, vol. 17,
no. 2, p. 273, 2017.
[125] Y. Wu, M. Yuan, S. Dong, L. Lin, and Y. Liu, “Remaining useful life
estimation of engineered systems using vanilla lstm neural networks,”
Neurocomputing, vol. 275, pp. 167–179, 2018.
[126] F. Wang, X. Fan, F. Wang, and J. Liu, “Backup battery analysis and allo-
cation against power outage for cellular base stations,” IEEE Transactions
on Mobile Computing, vol. 18, no. 3, pp. 520–533, 2018.
[127] S. Yi, Z. Hao, Z. Qin, and Q. Li, “Fog computing: Platform and applica-
tions,” in Proceedings of IEEE HotWeb, 2015, pp. 73–78.
[128] W. Shuai, P. Maillé, and A. Pelov, “Charging electric vehicles in the smart
city: A survey of economy-driven approaches,” IEEE Transactions on
Intelligent Transportation Systems, vol. 17, no. 8, pp. 2089–2106, 2016.
[129] J. Wang, Y. Ma, L. Zhang, R. X. Gao, and D. Wu, “Deep learning for
smart manufacturing: Methods and applications,” Journal of Manufactur-
ing Systems, vol. 48, pp. 144–156, 2018.
[130] F. Wang, F. Wang, X. Fan, and J. Liu, “Batalloc: Effective battery
allocation against power outage for cellular base stations,” in Proceedings
of ACM e-Energy, 2017, pp. 234–241.
[131] G. Kamath, P. Agnihotri, M. Valero, K. Sarker, and W.-Z. Song, “Pushing
analytics to the edge,” in Proceedings of IEEE GLOBECOM, 2016, pp.
1–6.
14 VOLUME 4, 2016

Deep Learning For Edge Computing Applications A ST

Uploaded by

Copyright:

Available Formats

Deep Learning For Edge Computing Applications A ST

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Deep Learning For Edge Computing Applications A ST

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in a future issue of this journal, but has not been

Deep Learning for Edge Computing

I. INTRODUCTION the cloud servers will carry out the computation-intensive

Edge Computing Deep Learning

Application Work Model Summary

Application Work Model Summary

Application Work Model Summary

Application Work Model Summary

or services to support edge-level learning and computing, 5 https://developer.apple.com/documentation/coreml?language=objc

You might also like