IEEE COMMUNICATIONS SURVEYS & TUTORIALS 1 Deep … · 2018-03-13 · IEEE COMMUNICATIONS SURVEYS & TUTORIALS 1 Deep Learning in Mobile and Wireless Networking: A Survey Chaoyun Zhang,

IEEE COMMUNICATIONS SURVEYS & TUTORIALS 1

Deep Learning in Mobile and Wireless Networking:A Survey

Chaoyun Zhang, Paul Patras, and Hamed Haddadi

Abstract—The rapid uptake of mobile devices and the risingpopularity of mobile applications and services pose unprece-dented demands on mobile and wireless networking infrastruc-ture. Upcoming 5G systems are evolving to support explodingmobile traffic volumes, agile management of network resourceto maximize user experience, and extraction of fine-grained real-time analytics. Fulfilling these tasks is challenging, as mobileenvironments are increasingly complex, heterogeneous, and evolv-ing. One potential solution is to resort to advanced machinelearning techniques to help managing the rise in data volumesand algorithm-driven applications. The recent success of deeplearning underpins new and powerful tools that tackle problemsin this space.

In this paper we bridge the gap between deep learningand mobile and wireless networking research, by presenting acomprehensive survey of the crossovers between the two areas.We first briefly introduce essential background and state-of-the-art in deep learning techniques with potential applications tonetworking. We then discuss several techniques and platformsthat facilitate the efficient deployment of deep learning ontomobile systems. Subsequently, we provide an encyclopedic reviewof mobile and wireless networking research based on deeplearning, which we categorize by different domains. Drawingfrom our experience, we discuss how to tailor deep learning tomobile environments. We complete this survey by pinpointingcurrent challenges and open future directions for research.

Index Terms—Deep Learning, Machine Learning, Mobile Net-working, Wireless Networking, Mobile Big Data, 5G Systems,Network Management.

I. INTRODUCTION

INTERNET connected mobile devices are penetrating everyaspect of individuals’ life, work, and entertainment. The

increasing number of smartphones and the emergence of ever-more diverse applications trigger a surge in mobile data traffic.Indeed, the latest industry forecasts indicate that the annualworldwide IP traffic consumption will reach 3.3 zettabytes(1015 MB) by 2021, with smartphone traffic exceeding PCtraffic by the same year [1]. Given the shift in user preferencetowards wireless connectivity, current mobile infrastructurefaces great capacity demands. In response to this increasing de-mand, early efforts propose to agilely provision resources [2]and tackle mobility management distributively [3]. In thelong run, however, Internet Service Providers (ISPs) must de-velop intelligent heterogeneous architectures and tools that canspawn the 5th generation of mobile systems (5G) and graduallymeet more stringent end-user application requirements [4], [5].

C. Zhang and P. Patras are with the Institute for Computing Systems Archi-tecture (ICSA), School of Informatics, University of Edinburgh, Edinburgh,UK. Emails: {chaoyun.zhang, paul.patras}@ed.ac.uk. H. Haddadi is with theDyson School of Design Engineering at Imperial College London. Email:[email protected].

The growing diversity and complexity of mobile networkarchitectures has made monitoring and managing the multi-tude of networks elements intractable. Therefore, embeddingversatile machine intelligence into future mobile networks isdrawing unparalleled research interest [6], [7]. This trend isreflected in machine learning (ML) based solutions to prob-lems ranging from radio access technology (RAT) selection [8]to malware detection [9], as well as the development ofnetworked systems that support machine leaning practices(e.g. [10], [11]). ML enables systematic mining of valuableinformation from traffic data and automatically uncover corre-lations that would otherwise have been too complex to extractby human experts [12]. As the flagship of machine learning,deep learning has achieved remarkable performance in areassuch as computer vision [13] and natural language processing(NLP) [14]. Networking researchers are also beginning torecognize the power and importance of deep learning, and areexploring its potential to solve problems specific to the mobilenetworking domain [15], [16].

Embedding deep learning into the 5G mobile and wirelessnetworks is well justified. In particular, data generated bymobile environments are heterogeneous as these are usuallycollected from various sources, have different formats, andexhibit complex correlations [17]. Traditional ML tools requireexpensive hand-crafted feature engineering to make accurateinferences and decisions based on such data. Deep learningeliminates domain expertise as it employs hierarchical featureextraction, by which it efficiently distills information andobtains increasingly abstract correlations from the data, whileminimizing the data pre-processing effort. Graphics ProcessingUnit (GPU)-based parallel computing further enables deeplearning to make inferences within milliseconds. This facil-itates network analysis and management with high accuracyand in a timely manner, overcoming the runtime limitations oftraditional mathematical techniques (e.g. convex optimization,game theory, meta heuristics).

Despite growing interest in deep learning in the mobilenetworking domain, existing contributions are scattered acrossdifferent research areas and a comprehensive yet concisesurvey is lacking. This article fills this gap between deeplearning and mobile and wireless networking, by presentingan up-to-date survey of research that lies at the intersectionbetween these two fields. Beyond reviewing the most rele-vant literature, we discuss the key pros and cons of variousdeep learning architectures, and outline deep learning modelselection strategies in view of solving mobile networkingproblems. We further investigate methods that tailor deeplearning to individual mobile networking tasks, to achieve best

arX

iv:1

803.

0431

1v1

[cs

.NI]

12

Mar

201

8


Related books,

surveys and

magazine papers

Our scope and

distinction

Overviews of

Deep Learning

Overviews of 5G

Network

Techniques and

Mobile Big Data

Crossovers

between Deep

Learning and

Mobile Networking

Evolution Principles AdvantagesMultilayer

Perceptron

Boltzmann

Machine

Auto-

encoder

Convolutional

Neural

Network

Recurrent

Neural

Network

Generative

Adversarial

Network

Deep Reinforcement

Learning

Model

Parallelism

Training

Parallelism

Serving Deep Learning with

Massive and High-Quality

Data

Unsupervised Deep Learning

Powered Analysis

Spatio-Temporal Mobile

Traffic Data Mining

Deep Reinforcement

Learning Powered Analysis

Advanced Parallel

Computing

Dedicated

Deep Learning

Libraries

Fog

Computing

Fast

Optimisation

Algorithms

Distributed Machine

Learning Systems

Tensorflow

Theano

(Py)Torch

Caffe(2)MXNET

Software

Hardware

Mobile Devices

and Systems

Distributed Data

Containers

Changing Mobile

Environment

Feature

Extraction

Big Data

Benefits

Multi-task

Learning

Unsupervised

Learning

Deep

Lifelong

Learning

Deep

Transfer

Learning

Deep Learning Driven

Mobile Data AnalysisDeep Learning Driven

Wireless Sensor Network

Deep Learning Driven

Network ControlDeep Learning Driven

Network SecurityEmerging Deep Learning Driven

Mobile Network Applications

Deep Learning Driven Mobility

Analysis and User Localisation

App-level Data

Analysis

Network Data

Analysis

Network

Prediction

Traffic

Classification

CDR

Mining

Mobile

HealthcareMobile Pattern

Recognition

Mobile NLP

and ASR

Anomaly

DetectionMalware

Detection

Botnet

DetectionPrivacy

Network

OptimisationRouting Scheduling

Resource

Allocation

Radio

ControlOthers

Network Data

Monetisation

IoT In-Network

Computation

Mobile

Crowdsensing

Mobility

Analysis

User

Localisation

Signal

Processing

Attacker

Sec. VI

Deep Learning Applications in Mobile & Wireless Network

Sec. VII

Tailored Deep Learning to Mobile Networks

Sec. VIII

Future Research Perspectives

Sec. II

Related Work & Our ScopeSec. III

Deep Learning 101

Sec. IV

Enabling Deep Learning in Mobile &

Wireless Networking

Sec. V

Deep Learning: State-of-the-Art

Fig. 1: Diagramatic view of the organization of this survey.

performance under complex environments. We wrap up thispaper by pinpointing future research directions and importantproblems that remain unsolved and are worth pursing withdeep neural networks. Our ultimate goal is to provide a definiteguide for networking researchers and practitioners, who intendto employ deep learning to solve problems of interest.

Survey Organization: We structure this article in a top-downmanner, as shown in Figure 1. We begin by discussing workthat gives a high-level overview of deep learning, future mobilenetworks, and networking applications built using deep learn-ing, which help define the scope and contributions of this paper(Section II). Since deep learning techniques are relatively newin the mobile networking community, we provide a basic deeplearning background in Section III, highlighting immediateadvantages in addressing mobile networking problems. Thereexist many factors that enable implementing deep learningfor mobile networking applications (including dedicated deeplearning libraries, optimization algorithms, etc.). We discussthese enablers in Section IV, aiming to help mobile networkresearchers and engineers in choosing the right software andhardware platforms for their deep learning deployments.

In Section V, we introduce and compare state-of-the-artdeep leaning models and provide guidelines for selectingamong candidates for solving networking problems. In Sec-tion VI we review recent deep learning applications to mobile

and wireless networking, which we group by different sce-narios ranging from mobile traffic analytics to security, andemerging applications. We then discuss how to tailor deeplearning models to mobile networking problems (Section VII)and highlight open issues related to deep learning adoption innetworking research (Section VIII). We conclude this articlewith a brief discussion on the interplay between mobilenetworks and deep neural networks (Section IX).1

II. RELATED HIGH-LEVEL ARTICLES ANDTHE SCOPE OF THIS SURVEY

Mobile networking and deep learning problems have beenresearched mostly independently. Only recently crossovers be-tween the two areas have emerged. Several notable works painta comprehensives picture of the deep learning or/and mobilenetworking research landscape. We categorize these works into(i) pure overviews of deep learning techniques, (ii) reviewsof analyses and management techniques in modern mobilenetworks, and (iii) reviews of works at the intersection betweendeep learning and computer networking. We summarize theseearlier efforts in Table II and in this section discuss the mostrepresentative publications in each class.

1We list the abbreviations used throughout this paper in Table I.


TABLE I: List of abbreviations in alphabetical order.

Acronym Explanation

5G 5th Generation mobile networksA3C Asynchronous Advantage Actor-Critic

AdaNet Adaptive learning of neural NetworkAE Auto-EncoderAI Artificial Intelligence

AMP Approximate Message PassingANN Artificial Neural NetworkASR Automatic Speech RecognitionBSC Base Station ControllerBP Back-Propagation

CDR Call Detail RecordCNN or ConvNet Convolutional Neural Network

ConvLSTM Convolutional Long Short-Term MemoryCPU Central Processing UnitCSI Channel State Information

CUDA Compute Unified Device ArchitecturecuDNN CUDA Deep Neural Network library

D2D Device to Device communicationDAE Denoising Auto-EncoderDBN Deep Belief Network

DenseNet Dense Convolutional NetworkDQN Deep Q-NetworkDRL Deep Reinforcement LearningDT Decision Tree

ELM Extreme Learning MachineGAN Generative Adversarial NetworkGP Gaussian Process

GPS Global Positioning SystemGPU Graphics Processing UnitGRU Gate Recurrent UnitHMM Hidden Markov ModelHTTP HyperText Transfer ProtocolIDS Intrusion Detection SystemIoT Internet of ThingsISP Internet Service ProviderLTE Long-Term Evolution

LS-GAN Loss-Sensitive Generative Adversarial NetworkLSTM Long Short-Term Memory

LSVRC Large Scale Visual Recognition ChallengeMDP Markov Decision ProcessMEC Mobile Edge ComputingML Machine Learning

MLP Multilayer PerceptronMIMO Multi-Input Multi-OutputMTSR Mobile Traffic Super-ResolutionNFL No Free Lunch theoremNLP Natural Language ProcessingNMT Neural Machine TranslationNPU Neural Processing Unitnn-X neural network neXtPCA Principal Components AnalysisQoE Quality of ExperienceRBM Restricted Boltzmann Machine

ResNet Residual NetworkRFID Radio Frequency IdentificationRNC Radio Network ControllerRNN Recurrent Neural Network

SARSA State-Action-Reward-State-ActionSGD Stochastic Gradient DescentSON Self-Organising NetworkSNR Signal-to-Noise RatioSVM Support Vector MachineTPU Tensor Processing UnitVAE Variational Auto-EncoderVR Virtual Reality

WGAN Wasserstein Generative Adversarial NetworkWSN Wireless Sensor Network

A. Overviews of Deep Learning and its Applications

The era of big data is triggering wide interest in deeplearning across different research disciplines [25]–[28] and agrowing number of surveys and tutorials are emerging (e.g.[21], [22]). LeCun et al. give a milestone overview of deeplearning, introduce several popular models, and look aheadat the potential of deep neural networks [18]. Schmidhuberundertakes an encyclopedic survey of deep learning, likelythe most comprehensive thus far, covering the evolution,methods, applications, and open research issues [19]. Liu et al.summarize the underlying principles of several deep learningmodels, and review deep learning developments in selectedapplications, such as speech processing, pattern recognition,and computer vision [20].

Arulkumaran et al. present several architectures and corealgorithms for deep reinforcement learning, including deep Q-networks, trust region policy optimization, and asynchronousadvantage actor-critic [24]. Their survey highlights the remark-able performance of deep neural networks in different controlproblem (e.g., video gaming, Go board game play, etc.). Zhanget al. survey developments in deep learning for recommendersystems [65], which have potential to play an important rolein mobile advertising. As deep learning becomes increasinglypopular, Goodfellow et al. provide a comprehensive tutorial ofdeep learning in a book that covers prerequisite knowledge,underlying principles, and popular applications [23].

B. Surveys on Future Mobile Networks

The emerging 5G mobile networks incorporate a host ofnew techniques to overcome the performance limitations ofcurrent deployments and meet new application requirements.Progress to date in this space has been summarized throughsurveys, tutorials, and magazine papers (e.g. [4], [5], [34], [35],[43]). Andrews et al. highlight the differences between 5G andprior mobile network architectures, conduct a comprehensivereview of 5G techniques, and discuss research challengesfacing future developments citeandrews2014will. Agiwal et al.review new architectures for 5G networks, survey emergingwireless technologies, and point out to research problems thatremain unsolved [4]. Gupta et al. also review existing workon 5G cellular network architectures, subsequently proposinga framework that incorporates networking ingredients such asDevice-to-Device (D2D) communication, small cells, cloudcomputing, and the IoT [5].

Intelligent mobile networking is becoming a popular re-search area and related work has been reviewed in the literature(e.g. [7], [30], [33], [48], [50]–[53]). Jiang et al. discussthe potential of applying machine learning to 5G networkapplications including massive MIMO and smart grids [7].This work further identifies several research gaps between MLand 5G that remain unexplored. Li et al. discuss opportunitiesand challenges of incorporating artificial intelligence (AI) intofuture network architectures and highlight the significance ofAI in the 5G era [52]. Klaine et al. present several successfulML practices in SONs, discuss the pros and cons of differentalgorithms, and identify future research directions in thisarea [51].


TABLE II: Summary of existing surveys, magazine papers, and books related to deep learning and mobile networking. Thesymbol D indicates a publication is in the scope of a domain; 7 marks papers that do not directly cover that area, but fromwhich readers may retrieve some related insights. Publications related to both deep learning and mobile networks are shaded.

Publication One-sentence summaryScope

Machine learning Mobile networkingDeep

leaningOther MLmethods

Mobilebig data

5G tech-nology

LeCun et al. [18] A milestone overview of deep learning. D

Schmidhuber [19] A comprehensive deep learning survey. D

Liu et al. [20] A survey on deep learning and its applications. D

Deng et al. [21] An overview of deep learning methods and applications. D

Deng [22] A tutorial on deep learning. D

Goodfellow et al. [23] An essential deep learning textbook. D 7

Arulkumaran et al. [24] A survey of deep reinforcement learning. D 7

Chen et al. [25] An introduction to deep learning for big data. D 7 7

Najafabadi [26] An overview of deep learning applications for big data analytics. D 7 7

Hordri et al. [27] A brief of survey of deep learning for big data applications. D 7 7

Gheisari et al. [28] A high-level literature review on deep learning for big data analytics. D 7

Yu et al. [29] A survey on networking big data. D

Alsheikh et al. [30] A survey on machine learning in wireless sensor networks. D D

Tsai et al. [31] A survey on data mining in IoT. D D

Cheng et al. [32] An introductions mobile big data its applications. D 7

Bkassiny et al. [33] A survey on machine learning in cognitive radios. D 7 7

Andrews et al. [34] An introduction and outlook of 5G networks. D

Gupta et al. [5] A survey of 5G architecture and technologies. D

Agiwal et al. [4] A survey of 5G mobile networking techniques. D

Panwar et al. [35] A survey of 5G networks features, research progress and open issues. D

Elijah et al. [36] A survey of 5G MIMO systems. D

Buzzi et al. [37] A survey of 5G energy-efficient techniques. D

Peng et al. [38] An overview of radio access networks in 5G. 7 D

Niu et al. [39] A survey of 5G millimeter wave communications. D

Wang et al. [2] 5G backhauling techniques and radio resource management. D

Giust et al. [3] An overview of 5G distributed mobility management. D

Foukas et al. [40] A survey and insights on network slicing in 5G. D

Taleb et al. [41] A survey on 5G edge architecture and orchestration. D

Mach and Becvar [42] A survey on MEC. D

Mao et al. [43] A survey on mobile edge computing. D D

Wang et al. [44] An architecture for personalized QoE management in 5G. D D

Han et al. [45] Insights to mobile cloud sensing, big data, and 5G. D D

Singh et al. [46] A survey on social networks over 5G. 7 D D

Chen et al. [47] An introduction to 5G cognitive systems for healthcare. 7 7 7 D

Buda et al. [48] Machine learning aided use cases and scenarios in 5G. D D D

Imran et al. [49] An introductions to big data analysis for self-organizing networks (SON) in 5G. D D D

Keshavamurthy et al. [50] Machine learning perspectives on SON in 5G. D D D

Klaine et al. [51] A survey of machine learning applications in SON. 7 D D D

Jiang et al. [7] Machine learning paradigms for 5G. 7 D D D

Li et al. [52] Insights into intelligent 5G. 7 D D D

Bui et al. [53] A survey of future mobile networks analysis and optimization. 7 D D D

Kasnesis et al. [54] Insights into employing deep learning for mobile data analysis. D D

Alsheikh et al. [17] Applying deep learning and Apache Spark for mobile data analytics. D D

Cheng et al. [55] Survey of mobile big data analysis and outlook. D D D 7

Wang and Jones [56] A survey of deep learning-driven network intrusion detection. D D D 7

Kato et al. [57] Proof-of-concept deep learning for network traffic control. D D

Zorzi et al. [58] An introduction to machine learning driven network optimisation. D D D

Fadlullah et al. [59] A comprehensive survey of deep learning for network traffic control. D D D 7

Zheng et al. [6] An introduction to big data-driven 5G optimisation. D D D D

Mohammadi et al. [60] A survey of deep learning in IoT data analytics. D D D

Ahad et al. [61] A survey of neural networks in wireless networks. D 7 7 D

Gharaibeh et al. [62] A survey of smart cities. D D D D

Lane et al. [63] An overview and introduction of deep learning-driven mobile sensing. D D D

Ota et al. [64] A survey of deep learning for mobile multimedia. D D D


C. Deep Learning Driven Networking Applications

A growing number of papers survey recent works thatbring deep learning into the computer networking domain.Alsheikh et al. identify benefits and challenges of using bigdata for mobile analytics and propose a Spark based deeplearning framework for this purpose [17]. Wang and Jonesdiscuss evaluation criteria, data streaming and deep learningpractices for network intrusion detection, pointing out researchchallenges inherent to such applications [56]. Zheng et al.put forward a big data-driven mobile network optimisationframework in 5G, to enhance QoE performance [6]. Morerecently, Fadlullah et al. deliver a survey on the progressof deep learning in a board range of areas, highlighting itspotential application to network traffic control systems [59].Their work also highlights several unsolved research issuesworthy of future study.

Ahad et al. introduce techniques, applications, and guide-lines on applying neural networks to wireless networkingproblems [61]. Despite several limitations of neural networksidentified, this article focuses largely on old neural networksmodels, ignoring recent progress in deep learning and success-ful applications in current mobile networks. Lane et al. inves-tigate the suitability and benefits of employing deep learningin mobile sensing, and emphasize on the potential for accurateinference on mobile devices [63]. Ota et al. report noveldeep learning applications in mobile multimedia. Their surveycovers state-of-the-art deep learning practices in mobile healthand wellbeing, mobile security, mobile ambient intelligence,language translation, and speech recognition. Mohammadi etal. survey recent deep learning techniques for Internet ofThings (IoT) data analytics [60]. They overview comprehen-sively existing efforts that incorporate deep learning into theIoT domain and shed light on current research challenges andfuture directions.

D. Our Scope

The objective of this survey is to provide a comprehensiveview on state-of-the-art deep learning practices in the mobilenetworking area. By this we aim to answer the following keyquestions:

1) Why is deep learning promising for solving mobilenetworking problems?

2) What are the cutting-edge deep learning models relevantto mobile and wireless networking?

3) What are the most recent successful deep learning appli-cations in the mobile networking domain?

4) How can researchers tailor deep learning to specificmobile networking problems?

5) Which are the most important and promising directionsworthy of further study?

The research papers and books we mentioned previouslyonly partially answer these questions. This article goes beyondthese previous works and specifically focuses on the crossoversbetween deep learning and mobile networking. While our mainscope remains the mobile networking domain, for complete-ness we also discuss deep learning applications to wireless net-works, and identify emerging application domains intimately

connected to these areas. Overall, our paper distinguishesitself from earlier surveys from the following perspectives:

(i) We particularly focus on deep learning applications formobile network analysis and management, instead ofbroadly discussing deep learning methods (as, e.g., in[18], [19]) or centering on a single application domain,e.g. mobile big data analysis with a specific platform [17].

(ii) We discuss cutting-edge deep learning techniques fromthe perspective of mobile networks (e.g., [66], [67]),focusing on their applicability to this area, whilst givingless attention to conventional deep learning models thatmay be out-of-date.

(iii) We analyze similarities between existing non-networkingproblems and those specific to mobile networks; basedon this analysis we provide insights into both best deeplearning architecture selection strategies and adaptationapproaches so as to exploit the characteristics of mobilenetworks for analysis and management tasks.

To the best of our knowledge, this is the first timethat mobile network analysis and management are jointlyreviewed from a deep learning angle. We also provide forthe first time insights into how to tailor deep learning tomobile networking problems.

III. DEEP LEARNING 101We begin with a brief introduction to deep learning, high-

lighting the basic principles behind computation techniquesin this field, as well as key advantages that lead to theirsuccess. Deep learning is essentially a sub-branch of MLwhere algorithms hierarchically extract knowledge from rawdata through multiple layers of nonlinear processing units,in order to make predictions or take actions according tosome target objective. The major benefit of deep learning overtraditional ML is automatic feature extraction, thus avoidingexpensive hand-crafted feature engineering. We illustrate therelation between deep learning, machine leaning, and artificialintelligence (AI) at a high level in Fig. 2.

AI

Machine Learning

Supervised learning

Unsupervised learning

Reinforcement learningDeep Learning

Examples:

MLP, CNN,

RNN

Applications in mobile &

wireless network

(our scope)

Examples: Examples:

Rule enginesExpert systems

Evolutionary algorithms

Fig. 2: Venn diagram of the relation between deep learning,machine learning, and AI. This survey particularly focuses ondeep learning applications in mobile and wireless networks.

The discipline traces its origins 75 years back, when thresh-old logic was employed to produce a computational model


for neural networks [68]. However, it was only in the late1980s that neural networks (NNs) gained interest, as Williamsand Hinton showed that multi-layer NNs could be trainedeffectively by back-propagating errors [69]. LeCun and Bengiosubsequently proposed the now popular Convolutional NeuralNetwork (CNN) architecture [70], but progress stalled due tocomputing power limitations of systems available at that time.Following the recent success of GPUs, CNNs have been em-ployed to dramatically reduce the error rate in the Large ScaleVisual Recognition Challenge (LSVRC) [71]. This has drawnunprecedented interest in deep learning and breakthroughscontinue to appear in a wide range of computer science areas.

A. Fundamental Principles of Deep Learning

The key aim of deep neural networks is to approximate com-plex functions through a composition of simple and predefinedoperations of units (or neurons). Such an objective functioncan be almost of any type, such as a mapping between imagesand their class labels (classification), computing future stockprices based on historical values (regression), or even decidingthe next optimal chess move given the current status on theboard (control). The operations performed are usually definedby a weighted combination of a specific group of hidden unitswith a non-linear activation function, depending on the struc-ture of the model. Such operations along with the output unitsare named “layers”. The neural network architecture resemblesthe perception process in a brain, where a specific set of unitsare activated given the current environment, influencing theoutput of the neural network model. In mathematical terms, thearchitecture of deep neural networks is usually differentiable,therefore the weights (or parameters) of the model can belearned my minimizing a loss function using gradient descentmethods through back-propagation, following the fundamentalchain rule [69]. We illustrate the principles of the learning andinference processes of a deep neural network in Fig. 3, wherewe use a CNN as example.

Inputs xOutputs y

Hidden

Layer 1

Hidden

Layer 2

Hidden

Layer 3

Forward Passing (Inference)

Backward Passing (Learning)

Units

Fig. 3: Illustration of the learning and inference processesof a 4-layer CNN. w(·) denote weights of each hidden layer,σ(·) is an activation function, λ refers to the learning rate,∗(·) denotes the convolution operation and L(w) is the lossfunction to be optimized.

B. Advantages of Deep Learning

We recognize several benefits of employing deep learningto address network engineering problems, namely:

1) It is widely acknowledged that, while vital to the perfor-mance of traditional ML algorithms, feature engineeringis costly [72]. A key advantage of deep learning isthat it can automatically extract high-level features fromdata that has complex structure and inner correlations.The learning process does not need to be designed bya human, which tremendously simplifies prior featurehandcrafting [18]. The importance of this is amplified inthe context of mobile networks, as mobile data is usuallygenerated by heterogeneous sources, is often noisy, andexhibits non-trivial spatial and temporal patterns [17],whose otherwise labeling would require outstanding hu-man effort.

2) Secondly, deep learning is capable of handling largeamounts of data. Mobile networks generate high volumesof different types of data at fast pace. Training traditionalML algorithms (e.g., Support Vector Machine (SVM) [73]and Gaussian Process (GP) [74]) sometimes requires tostore all the data in memory, which is computationallyinfeasible under big data scenarios. Furthermore, theperformance of ML does not grow significantly withlarge volumes of data and plateaus relatively fast [23]. Incontrast, Stochastic Gradient Descent (SGD) employedto train NNs only requires sub-sets of at each trainingstep, which guarantees deep learning’s scalability withbig data. Deep neural networks further benefit as trainingwith big data prevents model over-fitting.

3) Traditional supervised learning is only effective whensufficient labeled data is available. However, most cur-rent mobile systems generate unlabeled or semi-labeleddata [17]. Deep learning provides a variety of methodsthat allow exploiting unlabeled data to learn useful pat-terns in an unsupervised manner, e.g., Restricted Boltz-mann Machine (RBM) [75], Generative Adversarial Net-work (GAN) [76]. Applications include clustering [77],data distributions approximation [76], un/semi-supervisedlearning [78], [79], and one/zero shot learning [80], [81]amongst others.

4) Compressive representations learned by deep neural net-works can be shared across different tasks, while thisis limited or difficult to achieve in other ML paradigms(e.g., linear regression, random forest, etc.). Therefore, asingle model can be trained to fulfill multiple objectives,without requiring complete model retraining for differenttasks. We argue that this is essential for mobile networkengineering, as it reduces computational and memoryrequirements of mobile systems when performing multi-task learning applications [82].

Although deep learning can have unique advantages whenaddressing mobile network problems, it requires certain systemand software support, in order to be effectively deployed inmobile networks. We review and discuss such enablers in thenext section.


TABLE III: Summary of tools and techniques that enable deploying deep learning in mobile systems.

Technique Examples Scope Functionality Performanceimprovement

Energy con-sumption

Economiccost

Advancedparallel

computing

GPU, TPU [83],CUDA [84],cuDNN [85]

Mobile servers,workstations

Enable fast, paralleltraining/inference of deeplearning models in mobile

applications

High High Medium(hardware)

Dedicated deeplearning library

TensorFlow [86],Theano [87],Caffe [88],Torch [89]

Mobile serversand devices

High-level toolboxes thatenable network engineers tobuild purpose-specific deep

learning architectures

MediumAssociated

withhardware

Low(software)

Fog computingnn-X [90], ncnn [91],

Kirin 970 [92],Core ML [93]

Mobile devices Support edge-based deeplearning computing Medium Low Medium

(hardware)

Fast optimizationalgorithms

Nesterov [94],Adagrad [95],

RMSprop, Adam [96]

Training deeparchitectures

Accelerate and stabilize themodel optimization process Medium

Associatedwith

hardware

Low(software)

Distributedmachine learning

systems

MLbase [97],Gaia [10], Tux2 [11],

Adam [98],GeePS [99]

Distributed datacenters,

cross-server

Support deep learningframeworks in mobile

systems across data centersHigh High High

(hardware)

IV. ENABLING DEEP LEARNING IN MOBILE NETWORKING

5G systems seek to provide high throughput and ultra-lowlatency communication services, to improve users’ QoE [4].Implementing deep learning to build intelligence into 5Gsystems so as to meet these objectives is expensive, sincepowerful hardware and software is required to support train-ing and inference in complex settings. Fortunately, severaltools are emerging, which make deep learning in mobilenetworks tangible; namely, (i) advanced parallel computing,(ii) distributed machine learning systems, (iii) dedicated deeplearning libraries, (iv) fast optimization algorithms, and (v) fogcomputing. We summarize these advances in Table III andreview them in what follows.

A. Advanced Parallel Computing

Compared to traditional machine learning models, deepneural networks have significantly larger parameters spaces,intermediate outputs, and number of gradient values. Each ofthese need to be updated during every training step, requiringpowerful computation resources. The training and inferenceprocesses involve huge amounts of matrix multiplications andother operations, though they could be massively parallelized.Traditional Central Processing Units (CPUs) have limitednumber of cores thus they only support restricted computingparallelism. Employing CPUs for deep learning implementa-tions is highly inefficient and will not satisfy the low-latencyrequirements of mobile systems.

Engineers address this issues by exploiting the power ofGPUs. GPUs were originally designed for high performancevideo games and graphical rendering, but new techniques suchas Compute Unified Device Architecture (CUDA) [84] and theCUDA Deep Neural Network library (cuDNN) [85] developedby NVIDIA add flexibility to this type of hardware, allowingusers to customize their usage for specific purposes. GPUsusually incorporate thousand of cores and perform exception-ally in fast matrix multiplications required for training neuralnetworks. This provides higher memory bandwidth over CPUsand dramatically speeds up the learning process. Recent ad-vanced Tensor Processing Units (TPUs) developed by Google

even demonstrate 15-30× higher processing speeds and 30-80× higher performance-per-watt than CPUs and GPUs [83].

There are a number of toolboxes that can assist the com-putational optimization of deep learning on the server side.Spring and Shrivastava introduce a hashing based techniquethat substantially reduces computation requirements of deepnetworks implementations [100]. Mirhoseini et al. employ areinforcement learning scheme to enable machines to learnthe optimal operation placement over mixture hardware fordeep neural networks. Their solution achieves up to 20%faster computation speed than human experts’ designs of suchplacements [101].

Importantly, these systems are easy to deploy, thereforemobile network engineers do not need to rebuild mobileservers from scratch to support deep learning computing. Thismakes implementing deep learning in mobile systems feasibleand accelerates the processing of mobile data streams.

B. Distributed Machine Learning Systems

Mobile data is collected from heterogeneous sources (e.g.,mobile devices, network probes, etc.), and stored in multipledistributed data centers. With the increase of data volumes, itis impractical to move all mobile data to a central data centerto run deep learning applications [10]. Running network-widedeep leaning algorithms would therefore require distributedmachine learning systems that support different interfaces(e.g., operating systems, programming language, libraries), soas to enable training and evaluation of deep models acrossgeographically distributed servers simultaneously, with highefficiency and low overhead.

Deploying deep learning in a distributed fashion willinevitably introduce several system-level problems, whichrequire answering the following questions:

Consistency – When employing multiple machines to trainone model, how to guarantee model parameters and computa-tional processes are consistent across all servers?Fault tolerance – How to deal with equipment breakdownsin a large-scale distributed machine learning systems?Communication – How to optimize communication between


nodes in a cluster and how to avoid congestion?Storage – How to design efficient storage mechanisms tailoredto different environments (e.g., distributed clusters, singlemachines, GPUs), given I/O and data processing diversity?Resource management – How to assign workloads to nodesin a cluster, while making sure they work well-coordinated?Programming model – How to design programming inter-faces so that users can deploy machine learning, and how tosupport multiple programming languages?

There exist several distributed machine learning systemsthat facilitate deep learning in mobile networking applica-tions. Kraska et al. introduce a distributed system namedMLbase, which enables to intelligently specify, select, opti-mize, and parallelize ML algorithms [97]. Their system helpsnon-experts deploy a wide range of ML methods, allowingoptimization and running ML applications across differentservers. Hsieh et al. develop a geography-distributed MLsystem called Gaia, which breaks the throughput bottleneckby employing an advanced communication mechanism overWide Area Networks (WANs), while preserving the accuracyof ML algorithms [10]. Their proposal supports versatile MLinterfaces (e.g. TensorFlow, Caffe), without requiring signifi-cant changes to the ML algorithm itself. This system enablesdeployments of complex deep leaning applications over large-scale mobile networks.

Xing et al. develop a large-scale machine learning platformto support big data applications [102]. Their architectureachieves efficient model and data parallelization, enablingparameter state synchronization with low communication cost.Xiao et al. propose a distributed graph engine for ML namedTUX2, to support data layout optimization across machinesand reduce cross-machine communication [11]. They demon-strate remarkable performance in terms of runtime and con-vergence on a large dataset with up to 64 billion edges.Chilimbi et al. build a distributed, efficient, and scalablesystem named “Adam".2 tailored to the training of deepmodels [98] Their architecture demonstrates impressive perfor-mance in terms throughput, delay, and fault tolerance. Anotherdedicated distributed deep learning system called GeePS isdeveloped by Cui et al. [99]. Their framework allows dataparallelization on distributed GPUs, and demonstrates highertraining throughput and faster convergence rate.

C. Dedicated Deep Learning Libraries

Building a deep learning model from scratch can provecomplicated to engineers, as this requires definitions offorwarding behaviors and gradient propagation operationsat each layer, in addition to CUDA coding for GPUparallelization. With the growing popularity of deep learning,several dedicated libraries simplify this process. Most of thesetoolboxes work with multiple programming languages, andare built with GPU acceleration and automatic differentiationsupport. This eliminates the need of hand-crafted definitionof gradient propagation. We summarize these libraries below.

2Note that this is distinct from the Adam optimizer discussed in Sec. IV-D

TensorFlow3 is a machine learning library developed byGoogle [86]. It enables deploying computation graphs onCPUs, GPUs, and even mobile devices [103], allowing MLimplementation on both single and distributed architectures.Although originally designed for ML and deep neuralnetworks applications, TensorFlow is also suitable for otherdata-driven research purposes. Detailed documentation andtutorials for Python exist, while other programming languagessuch as C, Java, and Go are also supported. Building uponTensorFlow, several dedicated deep learning toolboxes werereleased to provide higher-level programming interfaces,including Keras4 and TensorLayer [104].

Theano is a Python library that allows to efficiently define,optimize, and evaluate numerical computations involvingmulti-dimensional data [87]. It provides both GPU andCPU modes, which enables user to tailor their programs toindividual machines. Theano has a large users group andsupport community, and was one of the most popular deeplearning tools, though its popularity is decreasing as coreideas and attributes are absorbed by TensorFlow.

Caffe(2) is a dedicated deep learning framework developedby Berkeley AI Research [88] and the latest version, Caffe2,5

was recently released by Facebook. It allows to train a neuralnetworks on multiple GPUs within distributed systems, andsupports deep learning implementations on mobile operationsystems, such as iOS and Android. Therefore, it has thepotential to play an important role in the future mobile edgecomputing.

(Py)Torch is a scientific computing framework with widesupport for machine learning models and algorithms [89]. Itwas originally developed in the Lua language, but developerslater released an improved Python version [105]. It is alightweight toolbox that can run on embedded systems suchas smart phones, but lacks comprehensive documentations.PyTorch is now officially supported and maintained byFacebook and mainly employed for research purposes.

MXNET is a flexible deep learning library that providesinterfaces for multiple languages (e.g., C++, Python, Matlab,R, etc.) [106]. It supports different levels of machine learningmodels, from logistic regression to GANs. MXNET providesfast numerical computation for both single machine and dis-tributed ecosystems. It wraps workflows commonly used indeep learning into high-level functions, such that standardneural networks can be easily constructed without substantialcoding effort. MXNET is the official deep learning frameworkin Amazon.

Although less popular, there are other excellent deep learn-ing libraries, such as CNTK,6 Deeplearning4j,7 and Lasagne,8

3TensorFlow,https://www.tensorflow.org/4Keras deep learning library, https://github.com/fchollet/keras5Caffe2, https://caffe2.ai/6MS Cognitive Toolkit, https://www.microsoft.com/en-us/cognitive-toolkit/7Deeplearning4j, http://deeplearning4j.org8Lasagne, https://github.com/Lasagne


which can be also employed in mobile systems. Selectingamong these varies according to specific applications.

D. Fast Optimization Algorithms

The objective functions to be optimized in deep learningare usually complex, as they include a sum of a extremelylarge number of data-wise likelihood functions. As the depthof the model increases, such functions usually exhibit highnon-convexity with multiple local minima, critical points, andsaddle points. In this case, conventional Stochastic GradientDescent (SGD) algorithms are slow in terms of convergence,which will restrict their applicability to latency constrainedmobile systems.

Over recent years, a set of new optimization algorithms havebeen proposed to speed up and stabilize the optimization pro-cess. Suskever et al. introduce a variant of the SGD optimizerwith Nesterov’s momentum, which evaluates gradients afterthe current velocity is applied [94]. Their method demonstratesfaster convergence rate when optimizing convex functions.Adagrad performs adaptive learning to model parameters ac-cording to their update frequency. This is suitable for handlingsparse data and significantly outperform SGD in terms ofrobustness [95]. RMSprop is a popular SGD based methodproposed by Hinton. RMSprop divides the learning rate byan exponential smoothing average of gradients and does notrequire one to set the learning rate for each training step [107].

Kingma and Ba propose an adaptive learning rate optimizernamed Adam that incorporates momentum by the first-ordermoment of the gradient [96]. This algorithm is fast in terms ofconvergence, highly robust to model structures, and is consid-ered as the first choice if one cannot decide which algorithmshould be used. Andrychowicz et al. suggest that the optimiza-tion process can be even learned dynamically [108]. They posethe gradient descent as a trainable learning problem, whichdemonstrates good generalization ability in neural networktraining. Wen et al. propose a training algorithm tailored todistributed systems [109]. They quantize float gradient valuesto {-1, 0 and +1} in the training processing, which theoreti-cally require 20 times less gradient communications betweennodes. The authors prove that such gradient approximationmechanism allows the objective function to converge to optimawith probability of 1, where in their experiments, only a 2%accuracy loss is observed on average on GoogleLeNet [110]training.

E. Fog Computing

The Fog computing paradigm presents a new opportunity toimplement deep learning in mobile systems. Fog computingrefers to a set of techniques that permit deploying applicationsor data storage at the edge of networks [111], e.g., onindividual mobile devices. This reduces the communicationsoverhead, offloads data traffic, reduces user-side latency, andlightens the sever-side computational burdens [112].

There exist several efforts that attempt to shift deep learn-ing computing from the cloud side to mobile devices. Forexample, Gokhale et al. develop a mobile coprocessor namedneural network neXt (nn-X), which accelerates the deep neural

networks execution in mobile devices, while retaining lowenergy consumption [90]. Bang et al. introduce a low-powerand programmable deep learning processor to deploy mo-bile intelligence on edge devices [113]. Their hardware onlyconsumes 288 µW but achieves 374 GOPS/W efficiency. ANeurosynaptic Chip called TrueNorth is proposed by IBM[114]. Their solution seeks to support computationally inten-sive applications on embedded battery-powered mobile de-vices. Qualcomm introduces a Snapdragon neural processingengine to enable deep learning computational optimizationtailored to mobile devices.9 Their hardware allows developersto execute neural network models on Snapdragon 820 boardsto serve a variety of applications. In close collaboration withGoogle, Movidius develops an embedded neural network com-puting framework that allows user-customized deep learningdeployments at the edge of mobile networks. Their productscan achieve satisfying runtime efficiency, while operatingwith ultra-low power requirements. More recently, Huaweiofficially announced the Kirin 970 as a mobile AI computingsystem on chip.10 Their innovative framework incorporatesdedicated Neural Processing Units (NPUs), which dramaticallyaccelerates neural network computing, enabling classificationof 2,000 images per second on mobile devices.

Beyond these hardware advances, there are also softwareplatforms that seek to optimize deep learning on mobiledevices (e.g., [115]). We compare and summarize all theseplatforms in Table IV.11 In additional to the mobile version ofTensorFlow and Caffe, Tencent released a lightweight, high-performance neural network inference framework tailored tomobile platforms, which relies on CPU computing.12 Thistoolbox performs better than all known CPU-based opensource frameworks in terms of inference speed. Apple hasdeveloped “Core ML", a private ML framework to facilitatemobile deep learning implementation on iOS 11.13 This lowersthe entry barrier for developers wishing to deploy ML modelson Apple equipment. Yao et al. develop a deep learningframework called DeepSense dedicated to mobile sensingrelated data processing, which provides a general machinelearning toolbox that accommodates a wide range of edgeapplications. It has moderate energy consumption and lowlatency, thus being amenable to deployment on smartphones.

The techniques and toolboxes mentioned above make thedeployment of deep learning practices in mobile networkapplications feasible. In what follows, we briefly introduceseveral representative deep learning architectures and discusstheir applicability to mobile networking problems.

9Qualcomm Helps Make Your Mobile Devices Smarter WithNew Snapdragon Machine Learning Software Development Kit:https://www.qualcomm.com/news/releases/2016/05/02/qualcomm-helps-make-your-mobile-devices-smarter-new-snapdragon-machine

10Huawei announces the Kirin 970- new flagship SoC with AI capabilitieshttp://www.androidauthority.com/huawei-announces-kirin-970-797788/

11Adapted from https://mp.weixin.qq.com/s/3gTp1kqkiGwdq5olrpOvKw12ncnn is a high-performance neural network inference framework opti-

mized for the mobile platform, https://github.com/Tencent/ncnn13Core ML: Integrate machine learning models into your app, https:

//developer.apple.com/documentation/coreml


TABLE IV: Comparison of mobile deep learning platform.

Platform Developer Mobile hardware supported Speed Code size Mobile compatibility Open-sourcedTensorFlow Google CPU Slow Medium Medium Yes

Caffe Facebook CPU Slow Large Medium Yesncnn Tencent CPU Medium Small Good Yes

CoreML Apple CPU/GPU Fast Small Only iOS 11+ supported NoDeepSense Yao et al. CPU Medium Unknown Medium No

V. DEEP LEARNING: STATE-OF-THE-ART

Revisiting Fig. 2, machine learning methods can be natu-rally categorized into three classes, namely supervised learn-ing, unsupervised learning, and reinforcement learning. Deeplearning architectures have achieved remarkable performancein all these areas. In this section, we introduce the keyprinciples underpinning several deep learning models anddiscuss their largely unexplored potential to solve mobilenetworking problems. We illustrate and summarize the mostsalient architectures that we present in Fig. 4 and Table V,respectively.

A. Multilayer Perceptron

The Multilayer Perceptrons (MLPs) is the initial ArtificialNeural Network (ANN) design, which consists of at leastthree layers of operations [130]. Units in each layer aredensely connected hence requiring to configure a substantialnumber of weights. We show an MLP with two hidden layersin Fig. 4(a). The MLP can be employed for supervised,unsupervised, and even reinforcement learning purposes. Al-though this structure was the most popular neural networkin the past, its popularity is decreasing as it entails highcomplexity (fully-connected structure), modest performance,and low convergence efficiency. MLPs are mostly used as abaseline or integrated into more complex architectures (e.g.,the final layer in CNNs used for classification). Building anMLP is straightforward, and it can be employed, e.g., to assistwith feature extraction in models built for specific objectives inmobile network applications. The advanced Adaptive learningof neural Network (AdaNet) enables MLPs to dynamicallytrain their structures to adapt to the input [116]. This newarchitecture can be potentially explored for analyzing contin-uously changing mobile environments.

B. Boltzmann Machine

Restricted Boltzmann Machines (RBMs) [75] were orig-inally designed for unsupervised learning purposes. Theyare essentially a type of energy-based undirected graphicalmodels, which include a visible layer and a hidden layer,and where each unit can only takes binary values (i.e., 0and 1). RBMs can be effectively trained using the contrastivedivergence algorithm [131] through multiple steps of Gibbssampling [132]. We illustrate the structure and the trainingprocess of an RBM in Fig. 4(b). RBM-based models areusually employed to initialize the weights of a neural networkin more recent applications. The pre-trained model can besubsequently fine-tuned for supervised learning purposes usinga standard BP algorithm. A stack of RBMs is called a Deep

Belief Network (DBN) [117], which performs layer-wise train-ing and achieves superior performance as compared to MLPsin many applications, including time series forecasting [133],ratio matching [134], and speech recognition [135]. Suchstructures can be even extended to a convolutional architecture,to learn hierarchical spatial representations [118].

C. Auto-Encoders

Auto-Encoders (AEs) are also designed for unsupervisedlearning and attempt to copy inputs to outputs. The under-lying principle of an AE is shown in Fig. 4(c). AEs arefrequently used to learn compact representation of data fordimension reduction [136]. Extended versions can be furtheremployed to initialize the weights of a deep architecture,e.g., the Denoizing Auto-Encoder (DAE) [119]), and generatevirtual examples from a target data distribution, e.g. VariationalAuto-Encoders (VAEs) [120]. AEs can be further employedto address network security problems, as several researchpapers confirm their effectiveness in detecting anomalies underdifferent circumstances [137]–[139], which we will furtherdiscuss in subsection VI-E.

The structures of RBMs and AEs are based upon MLPs,CNNs or RNNs. Their goals are similar, while their learningprocesses are different. Both can be exploited to extract pat-terns from unlabeled mobile data, which may be subsequentlyemployed for various supervised learning tasks, e.g., rout-ing [140], mobile activity recognition [141], periocular ver-ification [142] and base station user number prediction [143].

D. Convolutional Neural Networks

Instead of employing full connections between layers, Con-volutional Neural Networks (CNNs or ConvNets) employ aset of locally connected kernels (filters) to capture correla-tions between different data regions. This approach improvestraditional MLPs by leveraging three important ideas, namely,sparse interactions, parameter sharing, and equivariant repre-sentations [23]. This reduce the number of model parameterssignificantly and maintains the affine invariance (i.e., recogni-tion result are robust to the affine transformation of objects).We illustrate the operation of one 2D convolutional layer inFig. 4(d).

Owing to the properties mentioned above, CNNs achieveremarkable performance in imaging applications. Krizhevskyet al. [71] exploit a CNN to classify images on the Ima-geNet dataset [144]. Their method reduces the top-5 errorby 39.7% and revolutionizes the imaging classification field.GoogLeNet [110] and ResNet [121] significantly increase thedepth of CNN structures, and propose inception and residuallearning techniques to address problems such as over-fitting


Input layer

Hidden layers

Output layer

Feature 1

Feature 2

Feature 3

Feature 4

Feature 5

Feature 6

(a) Structure of an MLP with 2 hidden layers (blue circles).

Visible variables

Hidden variables

1. Sample from

2. Sample from

3. Update weights

(b) Graphical model and training process of an RBM. v and h denotevisible and hidden variables, respectively.

Input layer

Hidden

layer

Output layer

Minimise

(c) Operating principle of an auto-encoder, which seeks to reconstructthe input from the hidden layer.

(d) Operating principle of a convolutional layer.

...

1x 2x 3x tx

1S 2S 3S tS

th3h2h1h

Inputs

States

Outputs

(e) Recurrent layer – x1:t is the input sequence, indexed by time t, stdenotes the state vector and ht the hidden outputs. In a simple RNNlayer, the input may directly connect to its output without hidden states.

Output

Real data

Outputs from

the generator

OutputsInputs

Real/Fake

Generator Network

Discriminator Network

(f) Underlying principle of a generative adversarial network (GAN).

Outputs

Policy/Q values ...States Agent

EnvironmentActions

Rewards

Observed state

(g) Typical deep reinforcement learning architecture. The agent is a neural network model thatapproximates the required function.

Fig. 4: Typical structure and operation principles of MLP, RBM, AE, CNN, RNN, GAN, and DRL.


TABLE V: Summary of different deep learning architectures. GAN and DRL are shaded, since they are built upon other models.

Model Learningscenarios

Examplearchitectures

Suitableproblems Pros Cons

Potentialapplications in

mobile networks

MLPSupervised,

unsupervised,reinforcement

ANN,AdaNet [116]

Modeling datawith simplecorrelations

Naive structure andstraightforward to

build

High complexity,modest performance

and slowconvergence

Modelingmulti-attribute mobile

data; auxiliary orcomponent of otherdeep architectures

RBM UnsupervisedDBN [117],

ConvolutionalDBN [118]

Extracting robustrepresentation

Can generate virtualsamples Difficult to train well

Learningrepresentations from

unlabeled mobiledata; model weight

initialization;network flow

prediction

AE Unsupervised DAE [119],VAE [120]

Learning sparseand compact

representations

Powerful andeffective

unsupervised learning

Expensive to pretrainwith big data

model weightinitialization; mobile

data dimensionreduction; mobileanomaly detection

CNNSupervised,


AlexNet [71],ResNet [121],

3D-ConvNet [122],GoogleLeNet [110],

DenseNet [123]

Spatial datamodeling

Weight sharing;affine invariance

High computationalcost; challenging to

find optimalhyper-parameters;

requires deepstructures forcomplex tasks

spatial mobile dataanalysis

RNNSupervised,


LSTM [124],Attention based

RNN [125],ConvLSTM [126]

Sequential datamodeling

Expertise incapturing temporal

dependencies

High modelcomplexity; gradient

vanishing andexploding problems

Individual traffic flowanalysis;

network-wide(spatio-)temporal

data modeling

GAN Unsupervised WGAN [67],LS-GAN [127] Data generation

Can produce lifelikeartifacts from a target

distribution

Training process isunstable

(convergencedifficult)

Virtual mobile datageneration; assistingsupervised learning

tasks in network dataanalysis

DRL Reinforcement

DQN [128], deepPolicy

Gradient [129],A3C [66]

Control problemswith high-

dimensionalinputs

Ideal forhigh-dimensional

environmentmodeling

Slow in terms ofconvergence

Mobile networkcontrol and

management.

and gradient vanishing, introduced by “depth”. Their structureis further improved by the Dense Convolutional Network(DenseNet) [123], which reuses feature maps from each layer,thereby achieving significant accuracy improvements overother CNN based models, while requiring fewer layers. CNNhave been also extended to video applications. Ji et al. propose3D convolutional neural networks for video activity recog-nition [122], demonstrating superior accuracy as comparedto 2D CNN. More recent research focuses on learning theshape of convolutional kernels [145], [146]. These dynamicarchitectures allow to automatically focus on important regionsin input maps. Such properties are particularly important inanalyzing large-scale mobile environments exhibiting cluster-ing behaviors (e.g., surge of mobile traffic associated with apopular event).

Given the high similarity between image and spatial mobiledata (e.g., mobile traffic snapshots, users’ mobility, etc.),CNN-based models have huge potential for network-widemobile data analysis. This is a promising future direction thatwe further discuss in Sec. VIII.

E. Recurrent Neural NetworkRecurrent Neural Networks (RNNs) are designed for mod-

eling sequential data. At each time step, they produce out-

put via recurrent connections between hidden units [23],as shown in Fig. 4(e). Gradient vanishing and explodingproblems are frequently reported in traditional RNNs, whichmake them particularly hard to train [147]. The Long Short-Term Memory (LSTM) mitigates these issues by introducinga set of “gates” [124], which has been proven successfulin many applications (e.g., speech recognition [148], textcategorization [149], and wearable activity recognition [150]).Sutskever et al. introduce attention mechanisms to RNNmodels, which achieve outstanding accuracy in tokenizedpredictions [125]. Shi et al. substitute the dense matrix multi-plication in LSTMs with convolution operations, designing aConvolutional Long Short-Term Memory (ConvLSTM) [126].Their proposal reduces the complexity of traditional LSTMand demonstrates significantly lower prediction errors in pre-cipitation nowcasting.

Mobile networks produce massive sequential data fromvarious sources, such as data traffic flows, and the evolutionof mobile network subscribers’ trajectories and applicationlatencies. Exploring the RNN family is promising to enhancethe analysis of time series data in mobile networks.


F. Generative Adversarial Network

The Generative Adversarial Network (GAN) is a frameworkthat trains generative models using an adversarial process.It simultaneously trains two models: a generative one Gthat seeks to approximate the target data distribution fromtraining data, and a discriminative model D that estimatesthe probability that a sample comes from the real trainingdata rather than the output of G [76]. Both of G and D arenormally neural networks. The training procedure for G aimsto maximize the probability of D making a mistake. Each ofthem is trained iteratively while fixing the other one, finallyG can produce data close to a target distribution (the samewith training examples), if the model converges. We show theoverall structure of a GAN in Figure 4(f).

The training process of traditional GANs is highly sen-sitive to model structures, learning rates, and other hyper-parameters. Researchers are usually required to employ nu-merous ad hoc ‘tricks’ to achieve convergence. There existseveral solutions for mitigating this problem, e.g., WassersteinGenerative Adversarial Network (WGAN) [67] and Loss-Sensitive Generative Adversarial Network (LS-GAN) [127],but research on the theory of GANs remains shallow. Recentwork confirms that GANs can promote the performance ofsome supervised tasks (e.g., super-resolution [151], objectdetection [152], and face completion [153]) by minimizingthe divergence between inferred and real data distributions.Exploiting the unsupervised learning abilities of GANs ispromising in terms of generating synthetics mobile data forsimulations, or assisting specific supervised tasks in mobilenetwork applications. This becomes more important in taskswhere appropriate datasets are lacking, given that operatorsare generally reluctant to share their network data.

G. Deep Reinforcement Learning

Deep Reinforcement Learning (DRL) refers to a set ofmethods that approximate value functions (deep Q learning) orpolicy functions (policy gradient method) through deep neuralnetworks. An agent (neural network) continuously interactswith an environment and receives reward signals as feedback.The agent selects an action at each time step, which willchange the state of the environment. The training goal ofthe neural network is to optimize its parameters, such thatit can select actions that potentially lead to the best futurereturn. We illustrate this principle in Fig. 4(g). DRL is well-suited to problems that have a huge number of possible states(i.e., environments are high-dimensional). Representative DRLmethods include Deep Q-Networks (DQNs) [128], deep policygradient methods [129], and Asynchronous Advantage Actor-Critic [66]. These perform remarkably in AI gaming, robotics,and autonomous driving [154]–[157], and have made inspiringdeep learning breakthroughs over recent years.

On the other hand, many mobile networking problems canbe formulated as Markov Decision Processes (MDPs), wherereinforcement learning can play an important role (e.g., basestation on-off switching strategies [158], routing [159], andadaptive tracking control [160]). Some of these problemsnevertheless involve high-dimensional inputs, which limits the

applicability of traditional reinforcement learning algorithms.DRL techniques broaden the ability of traditional reinforce-ment learning algorithms to handle high dimensionality, inscenarios previously considered intractable. Employing DRLis thus promising to address network management and con-trol problems under complex, changeable, and heterogeneousmobile environments. We further discuss this potential inSec. VIII.

VI. DEEP LEARNING DRIVEN MOBILE AND WIRELESSNETWORKS

Deep learning has wide range of applications in mobilenetworking fields. We structure and group deep learningapplications in different networking domains and characterizetheir contributions. In what follows, we present the importantpublications across all areas and compare their design andprinciple.

A. Deep Learning Driven Mobile Data Analysis

The development of mobile technology (e.g. smartphones,augmented reality, etc.) are forcing mobile operators to evolvemobile network infrastructures. As a consequence, both cloudside and edge side of mobile networks are becoming increas-ingly sophisticated to cater the users’ requirements, whichfoster and consume huge amount of mobile data every day.These data can be either generated by sensors of mobiledevices that record individual users’ behaviors, or from mobilenetwork infrastructure which reflects the dynamics in urbanenvironments. Appropriately mining these data can benefitmultidisciplinary research fields, ranging from mobile networkmanagement and social analysis, to public transportation,personal services provision, and so on [32].

However, operators may be overwhelmed by managing andanalyzing massive amounts of heterogeneous mobile data.Deep learning is probably the most powerful methodologyfor the analysis of mobile big data. We begin this subsectionby introducing characteristics of mobile big data, then presenta holistic review on deep learning-driven mobile data analysis.

Characteristics of Mobile Data Yazti and Krishnaswamypropose to categorize mobile data into two groups, namelynetwork-level data and app-level data [161]. The key differencebetween them is that the former is usually collected by theedge mobile devices, while the latter one can be obtainedthrough network infrastructure. We summarize these two typesof data and their information comprised in Table VIII. Beforediscussing the mobile data analytics, we illustrate the datacollection process in Figure 5.

Mobile operators receive massive amount of network-levelmobile data generated by the networking infrastructure. Thesedata not only deliver a global view of mobile networkperformance (e.g. throughput, end-to-end delay, jitter, etc.),but also log individual session time, communication types,sender and receiver through Call Detail Records (CDRs). Thenetwork-level data usually exhibits significant spatio-temporalvariations resulting from users’ behaviors [162], which can


Base station

Cellular/WiFi network

BSC/RNC

Wireless sensor network

Sensor nodes

LocationTemperature

Humidity

Air quality

Noise

Weather

Reconnaissance

Magnetic field

Gateways

WiFi Modern

Sensor

data

Data

storage

App/Network level

mobile data

Mobile data

mining/analysis

Fig. 5: Illustration of the mobile data collection process incellular, WiFi and wireless sensor networks. BSC: Base

Station Controller; RNC: Radio Network Controller.

be utilized for network diagnosis and management, users’mobility analysis and public transportation planning [163].

On the other hand, App-level data is directly recordedby sensors or mobile applications installed in various mo-bile devices. These data is frequently collected throughcrowd-sourcing schemes from heterogeneous sources, such asGlobal Positioning Systems (GPS), mobile cameras and videorecorders, and potable medical monitors. Mobile devices act assensor hubs, which are responsible for data gathering and pre-processing, and subsequently distribute them to specific venuesas required [32]. App-level data usually directly or indirectlyreflects users’ behaviors, such as mobility, preference and so-cial links [55]. Analyzing app-level data from individuals canhelp reconstructing one’s personality and preference, whichcan be used for recommender systems and users targeting.The access to these data usually involves privacy and securityissues, which have to be addressed appropriately.

Mobile big data include several unique characteristics,such as spatio-temporal diversity, heterogeneity and personal-

TABLE VI: The taxonomy of mobile big data.

Mobile data Source Information

Network-level data

InfrastructureInfrastructure locations,capability, equipment

holders, etcPerformance

indicatorsData traffic, end-to-enddelay, QoE, jitter, etc.

Call detailrecords (CDR)

Session start and endtime, type, sender and

receiver, session results,etc.

Radioinformation

Signal power, frequency,spectrum, modulation etc.

App-level data

Device Device type, usage, MACaddress, etc.

Profile User setting, personalinformation, etc

SensorsMobility, temperature,

magnetic field,movement, etc

ApplicationPicture, video, voice,

health condition,preference, etc.

System log Log record software orhardware failure etc.

ity [32]. Some network-level data (e.g. mobile traffic) can beviewed as pictures taken by panoramic cameras, which providea city-scale sensing system for urban sensing. These imagescomprise information associated with movements of large-scale of individuals [162], hence exhibiting significant spatio-temporal diversity. Additionally, as modern smart phones canaccommodate multiple sensors and applications, one devicecan foster heterogeneous data simultaneously. Some of thesedata involve explicit identity information of individuals. Inap-propriate sharing and use can raise significant privacy issues.Therefore, extracting useful patterns from multi-source mobilewithout compromising user’s privacy remains a challengingendeavor.

Compared to traditional data analysis techniques, deeplearning embraces several unique features to address afore-mentioned challenges [17]. Namely:

1) Deep learning achieves remarkable performance in vari-ous data analysis tasks, on both structured and unstruc-tured data. Some types of mobile data can be representedas image-like (e.g. [163]) or sequential data [164].

2) Deep learning performs remarkably well in feature ex-traction from raw data. This saves tremendous effort ofhand-crafted feature engineering on mobile data, whichenables employees to spend more time on model design,and less sorting through the data itself.

3) Deep learning offer excellent tools (e.g. RBM, AE, GAN)for handing unlabeled data, which is common in mobilenetwork logs.

4) Multi-modal deep learning allows to learn features overmultiple modalities [165], which makes it powerful inmodeling to data collected from heterogeneous sensorsand data sources.

These advantages forge deep learning as a powerful tool formobile data analysis.


TABLE VII: A summary of work on network-level mobile data analysis.

Domain Reference Applications Model

Network prediction

Pierucci and Micheli [166] QoE prediction MLPGwon and Kung [167] Inferring Wi-Fi flow patterns Sparse coding + Max pooling

Nie et al. [168] Wireless mesh network traffic prediction DBN + Gaussian modelsMoyo and Sibanda [169] TCP/IP traffic prediction MLP

Wang et al. [170] Mobile traffic forecasting AE + LSTMZhang and Patras [171] Long-term mobile traffic forecasting ConvLSTM + 3D-CNN

Zhang et al. [163] Mobile traffic super-resolution CNN + GAN

Traffic classification

Wang [172] Traffic classification MLP, stacked AEWang et al. [173] Encrypted traffic classification CNN

Lotfollahi et al. [174] Encrypted traffic classification CNNWang et al. [175] Malware traffic classification CNN

CDR mining

Liang et al. [164] Metro density prediction RNNFelbo et al. [176] Demographics prediction CNNChen et al. [177] Tourists’ next visit location prediction MLP, RNNLin et al. [178] Human activity chains generation Input-Output HMM + LSTM

Others Xu et al. [179] Wi-Fi hotpot classification CNN

1) Network-Level Mobile Data Analysis: Network-level mobile data refers to logs recorded by Internetservice providers, including infrastructure metadata, networkperformance indicators and call detail records (CDRs) (seeTable. VIII). The recent remarkable success of deep learningignites global interests on exploiting this techniques formobile network-level data analysis, so as to optimize mobilenetworks configurations, thereby improving end-uses’ QoE.These work can be generally categorize into four subjectsaccording to applications, namely, network prediction,network traffic classification, CDR mining and radio analysis.In what follows, we review the success of deep learningachieves in these subjects. Before touching the details ofliteratures, we compare these work on Table VII for betterillustrations.

Network Prediction: Network prediction refers to inferringmobile network traffic or performance indicators given his-torical measurements or related data. Pierucci and Micheliinvestigate the relationship between key objective metrics andQoE [166]. They employ MLPs to predict users’ QoE inmobile communications given the average user throughput, thenumber of active users in the cells, the average data volumeper user, and the channel quality indicator, demonstrating highprediction accuracy. Network traffic forecasting is another fieldwhere deep learning is gaining importance. By leveragingsparse coding and max-pooling, Gwon and Kung developa semi-supervised deep learning model to classify receivedframe/packet patterns and infer the original properties offlows in a WiFi network [167]. Their proposal demonstratessuperior performance over traditional ML techniques. Nie etal. investigate the traffic demand patterns in wireless meshnetwork [168]. They design a DBN along with Gaussianmodels to precisely estimate traffic distributions.

In [170], Wang et al. propose to use an AE-based architec-ture and LSTMs to model spatial and temporal correlationsof mobile traffic distribution respectively. In particular, theirAE consists of a global stacked AE multiple local stackedAEs for spatial feature extraction, dimension reduction andtraining parallelism. Compressed representations extracted aresubsequently processed by LSTMs to perform final forecast-ing. Experiments on a real-world dataset demonstrate their

superior performance over SVM and Autoregressive IntegratedMoving Average model. An important work presented in [171]extend the mobile traffic forecasting to long term. The authorscombine ConvLSTM and 3D CNN to construct a spatio-temporal neural networks to capture the complex features theMilan city. They further introduce a fine-tune scheme andlightweight approach that blends prediction with historicalmean, which significantly extend the reliable prediction steps.

More recently, Zhang et al. propose an original MobileTraffic Super-Resolution (MTSR) technique to infer network-wide fine-grained mobile traffic consumption given coarse-grained counterparts obtained by probing, thereby reducingtraffic measurement overheads [163]. Inspired by image super-resolution techniques, they design a deep zipper network alongwith a Generative Adversarial Network (GAN) to performprecise MTSR and improve the fidelity of inferred trafficsnapshots. Their experiments over a real-world dataset showthat their architecture can improve the granularity of mobiletraffic snapshot by up to 100×, meanwhile significantly out-performing other interpolation techniques.Traffic Classification: The task of traffic classification is toidentify specific applications or protocols among the trafficin networks. Wang recognizes the powerful feature learningability of deep neural networks [172]. They use deep AEto identify protocols on a TCP flow dataset and achieveexcellent precision and recall rates. Work in [173] proposes touse 1D-CNN for encrypted traffic classification. They suggestthat 1D-CNN works well for modeling sequential data andhas lower complexity, thus being promising in addressingthe traffic classification problem. Similarly, Lotfollahi et al.present Deep Packet based on CNN for encrypted trafficclassification [174]. Their framework reduces the amountof hand-crafted feature engineering and achieves greataccuracy. CNN has also been used to identify malware traffic,where work in [175] regards traffic data as images andunusual patterns that malware traffic exhibit are classifiedby representation learning. Similar work on mobile malwaredetection will be further discussed in subsection VI-E.

CDR Mining: Telecommunications equipment is fosteringmassive CDR data everyday. This describes specific instancesof telecommunication transactions such as phone number, cell


ID, session start/end time, traffic consumption, etc. Usingdeep learning to mine useful information from CDR datacan serve a variety of functions. For example, Liang et al.propose Mercury to estimate metro density prediction fromstreaming CDR data using RNN [164]. They take trajectoryof a mobile phone user as a sequence of locations, while RNN-based models work well in handling sequential data. Likewise,Felbo et al. use CDR data to study demographics [176]. Theyemploy CNN to predict age and gender of mobile users,which demonstrates superior accuracy over other ML tools.More recently, Chen et al. compare different ML models topredict tourists’ next locations of visit via analyzing CDR data[177]. Their experiments suggest that RNN-based predictorsignificantly outperforms traditional ML methods, includingNaive Bayes, SVM, RF and MLP.

2) App-Level Mobile Data Analysis: Triggered by theincreasing popularity of Internet of Things (IoT), currentmobile devices install increasing number of applications andsensors that can collect massive amounts of app-level mo-bile data [180]. Employing artificial intelligence to extractuseful information from these data can extend the capabilityof devices [64], [181], [182], thus greatly benefiting usersthemselves, mobile operators as well as device manufacturers.This becomes an important and popular research direction inthe mobile networking domain. Nonetheless, mobile devicesusually operate in noisy, uncertain and unstable environments,where their users move fast and change their underlyinglocation and activity contexts frequently. As a result, adoptingtraditional machine learning tools to usually does not workwell app-level mobile data analysis. Advanced deep learningpractices provide a powerful solution for app-level data min-ing, as they demonstrate better precision and higher robustnessin IoT applications [183].

There exist two approaches for app-level mobile data anal-ysis, namely (i) cloud-based computing and (ii) edge-basedcomputing. We illustrate the difference between these scenar-ios in Fig. 6. The cloud-based computing treats mobile devicesas data collectors that constantly send data to servers withlimited local data preprocessing. Servers gather data receivedfor model training and inference, subsequently sending resultsback to each device (or locally store the analyzed results with-out dissemination, depending on application requirements).The drawback of this scenario is that users have to accessto the Internet so as to send or receive messages from servers,which will generate extra data transmission overhead and mayresult in severe latency for edge applications. The edge-basedcomputing scenario refers to offloading pre-trained modelsfrom the cloud to individual mobile device such that theycan make inferences locally. While this scenario requires lessinteractions with servers, its applicability is limited by thecapability of hardware and batteries. Therefore, it can onlysupport tasks that require light computations.

Many researchers employ deep learning for app-levelmobile data analysis. We group these work according totheir application domains, namely mobile healthcare, mobile

14Human profile source: https://lekeart.deviantart.com/art/male-body-profile-251793336

pattern recognition and mobile Natural Language Processing(NLP) and Automatic Speech Recognition (ASR). Table VIIIsummarizes the works in high level and we will discussseveral representative work next.

Mobile Health There are an increasing variety of wearablehealth monitoring devices being introduced to the market. Byequipping medical sensors, these devices can timely capturebody conditions of their carriers, providing real-time feedbacks(e.g. heart rate, blood pressure, breath status etc.), or triggeringalarms to remind users of taking medical actions [235]. Thisallows to provide timely and in-depth report to patiences anddoctors.

Liu and Du design a deep learning-driven MobiEar to aiddeaf people be aware of emergencies [184]. Their proposal ac-cepts acoustic signals as input, allowing users to register differ-ent acoustic events of interest. MobiEar operates efficiently onsmart phone and only requires infrequently communicationswith servers for update. Likewise, Liu et al. develop a UbiEaroperated on Android platform to assist hard-to-hear sufferersto recognize acoustic events without requirements of locationinformation [185]. Their design adopts a lightweight CNNarchitecture for inference acceleration while demonstratingcomparable accuracy over traditional CNN models.

Hosseini et al. design an edge computing system for healthmonitoring and treatment [190]. They use CNNs to extractfeatures from mobile sensor data, which plays an importantrole in their epileptogenicity localization application. Stamateet al. develop a mobile Android app called cloudUPDRS tomanage Parkinson’s symptoms for patiences [191]. In theirwork, MLPs are employed to determine acceptance of datacollected by smart phones to maintain high-quality data sam-ples. Their method proposed outperforms other ML methodssuch as GPs and RFs. Quisel et al. suggest that deep learningcan be effectively exploited for mobile health data analysis[192]. They exploit CNNs and RNNs to classify lifestyle andenvironmental traits of volunteers. Their models demonstratesuperior prediction accuracy over RFs and logistic regressionon six datasets.

As deep learning performs remarkably in medical dataanalysis [236], we expect more and more deep learning drivenhealth care devices emerge to help better physical monitoringand illness diagnosis.

Mobile Pattern Recognition Recent advanced mobile devicesoffer people a potable intelligent assistance, which fosters adiverse set of applications that can classify surrounding objects(e.g. [194]–[196], [199], [200], [207]) or users’ behaviors (e.g.[150], [203], [206], [212], [213], [237], [238]) based on mobilecamera or sensors. We review and compare recent works onmobile pattern recognition in this part.

Object classification in pictures taken by mobile devicesis drawing increasing research interest. Li et al. developDeepCham as a mobile object recognition framework [194].Their architecture involves a crowd-sourcing labeling processto reduce hand-labeling effort and a collaborative traininginstance generation pipeline tailored to deployment on mobiledevices. Evaluations on the prototype system suggest that their


Communications

Cloud-based

Data Collection

Model Training

& Updates

Inference

Results

Results

Storage and

analysis

Communications

Edge-based

Data Collection

Model

Pretraining

Local

Inference

Neural Network

Neural Network

Query

Results

Model

update

Fig. 6: Illustration of two deployment approaches for app-level mobile data analysis, namely cloud-based (left) andedge-based (right). The cloud-based approach makes inference on clouds and send results to edge devices. On the contrary,

the edge-based approach deploys models on edge devices which can make local inference.14

framework is efficient and effective in terms of training andinference. Tobías et al. investigate the applicability employingCNN schemes on mobile devices for objection recognitiontasks [195]. They conduct experiments on 3 deploymentscenarios (i.e. deployment (GPU, CPU and mobile) on 2benchmark datasets, which suggest that deep learning modelscan be efficiently embedded in mobile devices to perform real-time inference.

Mobile classifier can also assist applications on VirtualReality (VR). A CNN framework is proposed in [199]for facial expressions recognition when users are wearinghead-mounted displays in VR environment. They suggestthat this system can assist a real-time users interaction mode.Rao et al. incorporate a deep learning object detector into amobile augmented reality (AR) system [201]. Their systemachieves outstanding performance in detecting and enhancinggeographic objects in outdoor environments. Another workfocuses on mobile AR applications is introduced in [239],where the authors characterize of the tradeoffs betweenaccuracy, latency, and energy efficiency of object detection.

Activity recognition is another interesting field which relieson data collected by mobile motion sensors [238], [240].Heterogeneous sensors (e.g. motion, accelerometer, infrared,etc.) equipped on mobile devices can collect versatile typesof data that capture different activity information. Data col-lected will be delivered to servers for model training and themodel will be subsequently deployed to for domain-specifictasks. Essential features of sensor data can be automaticallyextracted by neural networks, making deep learning powerfulin performing complex activity recognition.

The first work based on deep learning appeared in 2014,where Zeng et al. employ a CNN to capture local dependencyand preserve scale invariance of motion sensor data [203].

They evaluate their proposal on 3 offline datasets, wheretheir proposal yields higher accuracy over statistics methodsand Principal components Analysis (PCA). Almaslukh et al.employ a deep AE to perform human activity recognitionby analyzing an offline smart phone dataset collected byaccelerometers and gyroscope sensors [204]. Li et al. considerdifferent scenarios for activity recognition [205]. In their im-plementation, Radio Frequency IDentification (RFID) data isdirectly sent to a CNN model for recognizing human activities.While their mechanism achieves high accuracies in differentapplications, experiments suggest that RFID-based methoddoes not work well for metal objects or liquid containers.

[206] exploits RBM to predict human activities given 7types of sensor data collected by smart watch. Experimentson prototype devices can efficiently fulfill the recognitionobjective under tolerable power requirements. Ordóñezand Roggen architect an advanced ConvLSTM to fusedata gathered from multiple sensors and perform activityrecognition [150]. By leveraging structures of CNN andLSTM, ConvLSTM can automatically compress spatio-temporal sensor data into low-dimensional representationswithout heavy data post-processing effort. Wang et al. exploitGoogle Soli to architect a mobile user-machine interactionplatform [208]. By analyzing radio frequency captured bymillimeter-wave radars, their architecture is able to recognize11 types of gestures with high accuracy. Their models aretrained on the server side, while making inferences locally onmobile devices.

Mobile NLP and ASR Recent remarkable achievementsobtained by deep learning on Natural Language Processing(NLP) and Automatic Speech Recognition (ASR) are shifttheir applications to mobile devices.

Powered by deep learning, the intelligent assistance Siri


TABLE VIII: A summary of works on app-level mobile data analysis.

Subject Reference Application Deployment Model

Mobile Healthcare

Liu and Du [184] Mobile ear Edge-based CNNLiu at al. [185] Mobile ear Edge-based CNN

Jindal [186] Heart rate prediction Cloud-based DBNKim et al. [187] Cytopathology classification Cloud-based CNN

Sathyanarayana et al. [188] Sleep quality prediction Cloud-based MLP, CNN, LSTMLi and Trocan [189] Health conditions analysis Cloud-based Stacked AEHosseini et al. [190] Epileptogenicity localisation Cloud-based CNNStamate et al. [191] Parkinson’s symptoms management Cloud-based MLPQuisel et al. [192] Mobile health data analysis Cloud-based CNN, RNNKhan et al. [193] Respiration surveillance Cloud-based CNN

Mobile Pattern Recognition

Li et al. [194] Mobile object recognition Edge-based CNN

Tobías et al. [195] Mobile object recognition Edge-based &Cloud based CNN

Pouladzadeh andShirmohammadi [196] Food recognition system Cloud-based CNN

Tanno et al. [197] Food recognition system Edge-based CNNKuhad et al. [198] Food recognition system Cloud-based MLP

Teng and Yang [199] Facial recognition Cloud-based CNNWu et al. [200] Mobile visual search Edge-based CNNRao et al. [201] Mobile augmented reality Edge-based CNN

Ohara et al. [202] WiFi-driven indoor change detection Cloud-based CNN,LSTMZeng et al. [203] Activity recognition Cloud-based CNN, RBM

Almaslukh et al. [204] Activity recognition Cloud-based AELi et al. [205] RFID-based activity recognition Cloud-based CNN

Bhattacharya and Lane [206] Smart watch-based activityrecognition Edge-based RBM

Antreas and Angelov [207] Mobile surveillance system Edge-based &Cloud based CNN

Ordóñez and Roggen [150] Activity recognition Cloud-based ConvLSTMWang et al. [208] Gesture recognition Edge-based CNN, RNNGao et al. [209] Eating detection Cloud-based DBM, MLPZhu et al. [210] User energy expenditure estimation Cloud-based CNN, MLP

Sundsøy et al. [211] Individual income classification Cloud-based MLPChen and Xue [212] Activity recognition Cloud-based CNNHa and Choi [213] Activity recognition Cloud-based CNN

Edel and Köppe [214] Activity recognition Edge-based Binarized-LSTM

Okita and Inoue [215] Multiple overlapping activitiesrecognition Cloud-based CNN+LSTM

Alsheikh et al. [17] Activity recognition using ApacheSpark Cloud-based MLP

Mittal et al. [216] Garbage detection Edge-based &Cloud based CNN

Seidenari et al. [217] Artwork detection and retrieval Edge-based CNNZeng et al. [218] Mobile pill classification Edge-based CNN

Lane and Georgiev [63] Mobile activity recognition, emotionrecognition and speaker identification Edge-based MLP

Yao et al. [219]Car tracking,heterogeneous human

activity recognition and useridentification

Edge-based CNN, RNN

Zeng [220] Mobile object recognition Edge-based UnknownKatevas et al. [221] Notification attendance prediction Edge-based RNN

Radu et al. [141] Activity recognition Edge-based RBM, CNNWang et al. [222], [223] Activity and gesture recognition Cloud-based Stacked AE

Cao et al. [224] Mood detection Cloud-based GRU

Mobile NLP and ASR

Wu et al. [225] Google’s Neural Machine Translation Cloud-based LSTM

Siri 15 Speech synthesis Edge-based Mixture densitynetworks

McGraw et al. [226] Personalised speech recognition Edge-based LSTMPrabhavalkar et al. [227] Embedded speech recognition Edge-based LSTM

Yoshioka et al. [228] Mobile speech recognition Cloud-based CNNRuan et al. [229] Shifting from typing to speech Cloud-based Unknown

Georgiev et al. [82] Multi-task mobile audio sensing Edge-based MLP

Others

Ignatov et al. [230] Mobile images quality enhancement Cloud-based CNN

Lu et al. [231] Information retrieval from videos inwireless network Cloud-based CNN

Lee et al. [232] Reducing distraction for smartwatchusers Cloud-based MLP

Vu et al. [233] Transportation mode detection Cloud-based RNNFang et al. [234] Transportation mode detection Cloud-based MLP


developed by Apple employs deep mixture density networks[241] fix its robotic voice [242]. This enables to synthesis amore human-like voice which sounds more natural to humans.An Android app released by Google is developed to supportmobile personalized speech recognition [226]. It quantizesparameters in LSTM model compression, allowing the app torun on low-power mobile phones. Likewise, Prabhavalkar etal. propose a mathematical RNN compression technique thatreduces the two third size of a LSTM acoustic model whileonly compromising negligible accuracy [227]. This allowsbuilding both memory- and energy-efficient ASR applicationson mobile devices.

Yoshioka et al. present their deep learning proposalsubmitted CHiME-3 challenge [243]. Their frameworkincorporates a network-in-network architecture into a CNNmodel which allows to perform ASR for mobile multi-microphone devices used under noisy environments [228].Mobile ASR can also accelerate the text input on mobiledevices, where Ruan et al.’s study shows that with the helpof ASR, input rates of English Mandarin are 3.0 and 2.8times faster over typing on keyboards [229]. They conductexperiments on a test-bed app based on iOS system, anddemonstrate impressive accuracy and acceleration by shiftingfrom typing to speech on mobile devices.

Others applications: Beyond aforementioned subjects, deeplearning also plays an important role in other applicationsin app-level data analysis. For instance, Ignatov et al. showthat deep learning can enhance the quality of pictures takenby mobile phones. By employing a CNN, they successfullyimprove the quality images obtained by different mobiledevices to a digital single-lens reflex camera level [230]. Luet al. focus on video post-processing under wireless networks[231], where their framework exploits a customized AlexNetto answer queries about object detections. This frameworkfurther involves an optimizer, which instructs mobile devicesto offload videos to reduce query response time.

Another interesting application is presented in [232], whereLee et al. show that deep learning can help smartwatch usersreduce distraction by eliminating unnecessary notifications.Specifically, they use an 11-layer MLP to predict the impor-tance of a notification. Fang et al. exploit a MLP to extractfeatures from high-dimensional and heterogeneous sensor data[234], which achieves high transportation mode recognitionaccuracy.

B. Deep Learning Driven Mobility Analysis and User Local-ization

Understanding movement patterns of a group of humanbeings is becoming crucial for epidemiology, urban planning,public service provisioning and mobile network resourcemanagement [244]. Location-based services and applications(e.g. mobile AR, GPS) demand precise individual positioningtechnology [245]. As a result, research on user localizationis evolving rapidly, cultivating a set of emerging positioningtechniques [246]. In this subsection, we give an overview onrelated work in Table IX and discuss them next.

Mobility Analysis: Since deep learning is able to capture spa-tial dependency in sequential data, it is becoming a powerfultool for mobility analysis. The applicability of deep learningfor trajectory prediction is studied in [248]. By sharing repre-sentations learned by RNN and Gate Recurrent Unit (GRU),the framework can perform multi-task learning on both socialnetworks and mobile trajectories modeling. Specifically, theyfirst use deep learning to reconstruct social network representa-tions of users, subsequently employing RNN and GRU modelsto learn patterns of mobile trajectories with different timegranularity. Importantly, these two components jointly sharerepresentations learned, which tightens the overall architectureand enables efficient implementation. Ouyang et al. arguemobility data are normally high-dimensional which may beproblematic for traditional ML models. Instead they build upondeep learning advances and propose a online learning schemeto train a hierarchical CNN architecture, allowing modelparallelism for data stream processing [247]. By analyzingusage detail records, their framework “DeepSpace” predictsindividuals’ trajectories with much higher accuracy on a real-world dataset as compared to naive CNNs.

Instead of focusing on individual trajectories, Song et al.shed light on the mobility analysis on a larger scale [249]. Intheir work, LSTM networks are exploited to jointly model thecity-wide movement patterns of a large group of people andvehicles. Their multi-task architecture demonstrates superiorprediction accuracy over vanilla LSTM. City-wide mobilepatterns is also researched in [250], where the authors archi-tect a deep spatio-temporal residual networks to forecast themovement of crowds. In order to capture the unique charac-teristics of spatio-temporal correlations associated with humanmobility, their framework abandons RNN-based models andconstructs three ResNets to extract nearby and distant spatialdependencies of human mobility in a city. This scheme learnstemporal features and fuses all representations extracted byall models for the final prediction. By incorporating externalevents information, their proposal achieves the highest accu-racy among all deep learning and non-deep learning methods.

Lin et al. consider generating human movement chainsfrom cellular data to support transportation planning [178]. Inparticular, they first employ an input-output Hidden MarkovModel (HMM) to label activity profiles for CDR data pre-processing. Subsequently, a LSTM is designed for activitychain generation, given the labeled activity sequences. Theyfurther synthesize urban mobility plans using the generativemodel and the simulation yields decent fit accuracy as thereal data.User Localization: Deep learning is also playing an importantrole in user localization. To overcome the variability andcoarse-grained limitations of signal strength based methods,Wang et al. propose a deep learning driven fingerprintingsystem name “DeepFi” to perform indoor localization based onChannel State Information (CSI) [253]. Their toolbox yieldsmuch higher accuracy than traditional methods including in-cluding FIFS [264], Horus [265], and Maximum Likelihood[266]. The same group of authors extend their work in [222],[223] and [254], [255], where they update the localizationsystem, such that it can work with calibrated phase information


TABLE IX: A summary of work on deep learning driven mobility analysis and indoor localization.

Domain Reference Application Model Key contribution

Mobility analysis

Ouyang et al. [247] Mobile user trajectoryprediction CNN Online framework for data stream

processing

Yang et al. [248] Social Networks & mobiletrajectories modeling RNN, GRU Multi-task learning

Song et al. [249] City-wide mobility prediction& transportation modelling Multi-task LSTM Multi-task learning

Zhang et al. [250] Citywide crowd flowsprediction

Deep spatio-temporalresidual networks

(CNN-based)

Exploitation of spatio-temporalcharacteristics of mobility external

events

Lin et al. [178] Human activity chainsgeneration

Input-Output HMM +LSTM Generative model

Subramanian and Sadiq[251] Mobile movement prediction MLP Require less location update and

paging signaling costs

Ezema and Ani [252] Mobile location estimation MLPOperate with received signal

strength in Global System forMobile Communications

User localization

Wang et al. [253] Indoor fingerprinting RBM First deep learning driven indoorlocalization based on CSI

Wang et al. [222], [223] Indoor localization RBM Work with calibrated phaseinformation of CSI

Wang et al. [254] Indoor localization CNN Use more robust angle of arrivingfor estimation

Wang et al. [255] Indoor localization RBMBi-modal framework using both

angle of arriving and averageamplitudes of CSI

Nowicki andWietrzykowski [256] Indoor localization Stacked AE Require less effort for system

tuning or filtering

Wang et al. [257], [258] Indoor localization Stacked AE Device-free framework, multi-tasklearning

Mohammadiet al. [259] Indoor localization VAE+DQNHandle unlabeled data;

reinforcement learning aidedsemi-supervised learning

Kumar et al. [260] Indoor vehicles localization CNN Focus on vehicles applications

Zheng and Weng [261] Outdoor navigation Developmental network Online learning scheme;edge-based

Zhang et al. [262] Indoor and outdoorlocalization Stacked AE Operate under both indoor and

outdoor environments

Vieira et al. [263] Massive MIMOfingerprint-based positioning CNN Operate with massive MIMO

channels

of CSI [222], [223]. They further use more sophisticated CNN[254] and bi-modal [255] to improve the accuracy.

Nowicki and Wietrzykowski [256] propose to exploit deeplearning to reduce effort of indoor localization. There frame-work obtains satisfactory prediction performance, while reduc-ing significantly effort for system tuning or filtering. Wanget al. suggest that the objective of indoor localization canbe achieved without the help of mobile devices. In [258],the authors employ an AE to learn useful patterns fromWiFi signals. By automatic feature extraction, they produce apredictor that can fulfill multi-tasks simultaneously, includingindoor localization, activity, and gesture recognition. Kumaret al. use deep learning to address the problem of indoorvehicles localization [260]. They employ CNNs to analyzevisual signal to localize vehicles in a car park. This can helpdriver assistance systems operate in underground environmentswhere the system has limited available vision.

Most mobile devices can only produce unlabeled positiondata, therefore unsupervised and semi-supervised learningbecome essential. Mohammadi et al. address this problem byleveraging of DRL and VAE. In particular, their framework en-visions a virtual agent in indoor environments [259]. This canconstantly receive state information during training, includingsignal strength indicators, current location of the agent and

the real (labeled data) and inferred (via a VAE) distance to thetarget. The agent can virtually move to eight directions at eachtime step. Each time it takes an action, it receives an rewardsignal, identifying whether it moves to a correct direction. Byemploying deep Q learning, the agent can finally accuratelylocalize user given both labeled and unlabeled data. Thiswork represents a important step towards IoT data analysisapplications, as it only requires limit supervision.

Beyond indoor localization, there also exist several researchworks that apply deep learning to the outdoor scenario. Forexample, Zheng and Weng introduce a lightweight develop-mental network for outdoor navigation application on mobiledevices [261]. Compared to CNNs, their architecture required100 times fewer weights to update compared to originalCNN while maintaining decent accuracy. This enables efficientoutdoor navigation implementation on mobile devices. Workin [262] studies localization under both indoor and outdoorenvironments. They use an AE to pre-train a four-layer MLP,in order to avoid hand-craft feature engineering. The MLP cansubsequently used to estimate the coarse position of targets.They further introduce an HMM to fine-tune the predictionbased on temporal properties of data. This improves theaccuracy estimation in both in/outdoor positioning with Wi-Fi signals.


TABLE X: A summary of work on deep learning drivenWSNs.

Reference Application ModelChuang and Jiang [267] Node localization MLP

Bernas and Płaczek [268] Indoor localization MLPPayal et al. [269] Node localization MLPDong et al. [270] Underwater localization MLP

Yan et al. [271] Smoldering and flamingcombustion identification MLP

Wang et al. [272] Temperature correction MLPLee et al. [273] Online query processing CNN

Li and Serpen [274] Self-adaptive WSN Hopfieldnetwork

Khorasani and Naji [275] Data aggregation MLPLi et al. [276] Distributed data mining MLP

C. Deep Learning Driven Wireless Sensor Networks

The Wireless Sensor Networks (WSNs) consists of a setof unique or heterogeneous sensors that are distributed overgeographical regions. Theses sensors collaboratively monitorphysical or environment status (e.g. temperature, pressure, mo-tion and pollution) and transmit data collected to centralizedservers through wireless channels, see wireless sensor network(purple circle in Fig. 5 for an illustration). A WSN typicallyinvolves three key core tasks, namely sensing, communicationand analysis. Deep learning is becoming increasingly popularfor WSN data analysis. In what follows, we review worksof deep learning driven WSN. Note that this is distinct frommobile data analysis we discuss at subsection VI-A, as inthis subsection we only focus on WSN applications. Beforestarting, we summaries these works in Table X.

There exist two data processing scenarios in WSNs, namelycentralized and decentralized. The former simply takes sensorsas data collectors, who are only responsible for gathering dataand send them to a central server for processing. The latternevertheless assumes sensors have somewhat computationalability. The main server offloads part of the jobs to the edgeand each sensor will perform data processing individually.Work in [275] focuses on the centralized scheme, where theauthors apply a 3-layer MLP to reduce data redundancy whilemaintaining essential points for data aggregation. These dataare sent to a central server for analysis. In contrast, Li et al.propose to distribute data mining to individual sensors [276].They partition a deep neural network into different layers andoffload layer operations to sensor nodes. Simulations con-ducted suggest that by pre-processing with neural networks,their framework obtains high fault detection accuracy, whilereducing power consumption at the central server.

Chuang and Jiang exploit neural networks to localize sensornodes in WSNs [267]. To adapt deep learning model to specificnetwork topology, they employ an online training schemeand correlated topology-trained data, enabling efficient modelsimplementations and accurate location estimations. Based onthis, Bernas and Płaczek architect an ensemble system thatinvolves multiple MLPs for location estimation in differentregions of interest [268]. In this scenario, node locationsinferred by multiple MLPs are fused by a fusion algorithm,which improves the localization accuracy, particularly bene-fiting to sensor nodes that are around boundaries of regions.

A comprehensive comparison of different training algorithmsapply MLP-based node localization is presented in [269].Their experiments suggest that the Bayesian regularizationalgorithm in general yields the best performance. Dong et al.consider an underwater node localization scenario [270]. Sinceacoustic signal suffers from the loss caused by absorption,scattering, noise, and interference, underwater localization isnot straightforward. By adopting a deep neural network, theirframework successfully addresses the aforementioned chal-lenges and achieves higher inference accuracy as comparedto SVM and generalized least square methods.

Deep learning has also been exploited for identificationof smoldering and flaming combustion phases in forests. In[271], Yan et al. embed a set of sensors into a forest tomonitor CO2, smoke, and temperature. They suggest thatvarious burning scenarios will emit different gases, whichcan be taken advantage to classify smoldering and flamingcombustion. Wang et al. consider to apply deep learning tocorrect inaccurate measurements of air temperature [272].They discover a close relationship between solar radiation andactual air temperature, which can be effectively learned byneural networks.

Missing data or de-synchronization are common in WSNdata collection. These may lead to serious problems in analysisdue to the inconsistency. Lee et al. address this problem byplugging in a query refinement component in a deep learningbased WSN analysis systems [273]. They employ exponentialsmoothing to infer missing data, thereby maintaining theintegrity of data for deep learning analysis without signif-icantly compromising accuracy. To enhance the intelligenceof WSNs, Li and Serpen embed an artificial neural networkinto a WSN, allowing it to agilely react to potential changesand following deployment in the field [274]. To this end,they employ a minimum weakly-connected dominating set torepresent the topology for the WSN, and subsequently use aHopfield recurrent neural network as a static optimizer to adaptnetwork infrastructure to potential changes as necessary. Thiswork represents an important step toward embedding machineintelligences to WSNs.

Turning attention again to Table X, it is interesting to seethat the majority of deep learning practices on WSN employMLP models. Since MLP is straightforward to architect andperform reasonably well, it remains a good candidate for WSNapplications. On the other hand, since most of sensor datacollected is sequential, we are expecting RNN-based modelscan play a more important role in this area. This can be apromising future research direction.

D. Deep Learning Driven Network Control

In this part, we turn our attention to mobile network controlproblems. Due to the powerful function approximation mecha-nism, deep learning has made remarkable breakthroughs in im-proving traditional reinforcement learning [24] and imitationlearning [277]. These advances have potential to solve mobilenetwork control problems which are previously complex orintractable [278], [279]. Recall that in reinforcement learning,an agent continuously interacts with the environment to learn


State

representation

Neural network agent

Actions

Rewards

Observed state

Mobile network

environment

Observation

representation

Neural network agent

Observations

Mobile network

environment

Demonstration

Predicted

actions

Desired

actions

Learning

Observation

representation

Neural network analysis

Observations

Mobile network

environment

Analysis

results

Outer

controller

Reinforcement Learning

Imitation Learning

Analysis-based Control

Fig. 7: Principles of three control approaches applied inmobile and wireless networks control, namely reinforcementlearning (above), imitation learning (middle), and analysis-based control (below).

the best action. With constant exploration and exploitation,the agent learns to perform without feedback to maximize itsexpected return. Imitation learning follows a different learningparadigm called “learning by demonstration”. This learningparadigm relies on a ‘teacher’ (usually a human) who tells theagent what action should be executed under certain observa-tions during the training. After sufficient demonstrations, theagent learns a policy that imitate the behavior of its teacherand can operate itself without supervision.

Beyond these two learning scenarios, analysis-based controlis gaining attraction in mobile networking. Specifically, thisscheme uses ML learning models for network data analysis,and subsequently exploits the results to aid network control.Unlike the prior scenarios, analysis-based control paradigmdoes not directly output the actions. Instead, it extract usefulinformation and deliver them to an extra agent to executethe actions. We illustrate the principles between the threecontrol paradigms in Fig. 7. We reviews works which havebeen proposed so far in this subsection and summarize themin Table XI.

Network Optimization: Network Optimization refers to the

management of network resources and functions for a givenenvironment to improve the network performance. Deeplearning recently achieved several successful results in thisarea. For example, Liu et al. exploit a DBN to discoverthe correlations between multi-commodity flow demandinformation and link usage in wireless networks [302]. Basedon the predictions, they remove the links that are unlikelyto be scheduled, so as to reduce the size of data for thedemand constrained energy minimization. Their methodreduces runtime by up to 50%, without compromising theoptimality. Subramanian and Banerjee propose to use deeplearning to predict the health condition of heterogeneousdevices in machine to machine communications [280]. Theresults obtained are subsequently exploited for optimizinghealth aware policy change decisions. He et al. employ deepreinforcement learning to address caching and interferencealignment problems in wireless networks [281], [282]. Inparticular, they treat time-varying channels as finite-stateMarkov channels and apply deep Q networks to learnthe best user selection policy in interference alignmentwireless networks. This novel framework demonstratessignificantly higher sum rate and energy efficiency overexisting approaches.

Routing: Deep learning can also improve the efficiency ofrouting rules. Lee et al. exploit a 3-layer deep neural networkto classify node degree given detailed information of therouting node [284]. The classification results along withtemporary routes are exploited for the following virtual routegeneration using the Viterbi algorithm. Mao et al. employDBN to decide the next routing node to construct a softwaredefined router [140]. By considering Open Shortest Path Firstas the optimal routing strategy, their method achieves upto 95% accuracy, while reducing significantly less overheadand delay, and achieving higher throughput with signalinginterval of 240 milliseconds. A similar conclusion is obtainedin [285], where the authors employ Hopfield neural networksfor routing, achieving better usability and survivability inmobile ad hoc network application scenarios.

Scheduling: There are several studies investigate schedul-ing with deep learning. Zhang et al. introduce a deep Qlearning-powered hybrid dynamic voltage and frequency scal-ing scheduling mechanism to reduce the energy consumptionin real-time systems (e.g. Wi-Fi communication, IoT, videoapplications) [287]. In their proposal, an AE is employed toapproximately the Q function, and the framework performexperience replay [303] to stabilize the training process andaccelerate the convergence. Simulations demonstrate that thismethod reduces by 4.2% energy consumption over a traditionalQ learning based method. Similarly, the work in [288] usesdeep Q learning for scheduling in roadside communicationnetworks. In particular, interactions between vehicular envi-ronments, including the sequence of actions, observations, andreward signals are formulated as a MDP. By approximatingthe Q value function, the agent learns a scheduling policythat achieves less incomplete requests, latency and busy timeand longer battery life time compared to traditional scheduling


TABLE XI: A summary of work on deep learning driven network control.

Domain Reference Application Control approach Model

Network optimization

Liu et al. Liu et al. Demand constrained energyminimization Analysis-based DBN

Subramanian andBanerjee [280]

Machine to machine systemoptimization Analysis-based Deep multi-modal

network

He et al. [281], [282] Caching and interferencealignment Reinforcement learning Deep Q learning

Masmar and Evans [283] mmWave Communicationperformance optimization Reinforcement learning Deep Q learning

Routing

Lee et al. [284] Virtual route assignment Analysis-based MLPYang et al. [285] Routing optimization Analysis-based Hopfield neural networksMao et al. [140] Software defined routing Imitation learning DBNTang et al. [286] Wireless network routing Imitation learning CNN

Scheduling

Zhang et al. [287] Hybrid dynamic voltage andfrequency scaling scheduling Reinforcement learning Deep Q learning

Atallah et al. [288] Roadside communication networkscheduling Reinforcement learning Deep Q learning

Chinchali et al. [289] Cellular network traffic scheduling Reinforcement learning Policy gradient

Atallah et al. [288] Roadside communicationnetworks scheduling Reinforcement learning Deep Q learning

Resource allocation

Sun et al. [290] Resource management overwireless networks Imitation learning MLP

Xu et al. [291] Resource allocation in cloud radioaccess networks Reinforcement learning Deep Q learning

Ferreira et al. [292] Resource management incognitive space communications Reinforcement learning

Deep State-Action-Reward-State-Action

(SARSA)

Ye and Li [293] Resource allocation invehicle-to-vehicle communication Reinforcement learning Deep Q learning

Radio control

Naparstek and Cohen[294] Dynamic spectrum access Reinforcement learning Deep Q learning

O’Shea and Clancy [295] Radio control and signal detection Reinforcement learning Deep Q learning

Wijaya et al. [296], [297] Intercell-interference cancellationand transmit power optimisation Imitation learning RBM

OthersMao et al. [298] Adaptive video bitrate Reinforcement learning A3C

Oda et al. [299], [300] Mobile actor node control Reinforcement learning Deep Q learningKim [301] IoT load balancing Analysis-based DBN

methods.

More recently, Chinchaliet al. present a policy gradientbased scheduler to optimize cellular network traffic flow[289]. Specifically, they cast the scheduling problem as aMDP and employ RF to predict network throughput whichis subsequently used as a component as a reward function.Evaluations over a realistic network simulator demonstratethat their proposal can dynamically adapt to traffic variation,which enables mobile networks to carry 14.7% more datatraffic while outperforming heuristic schedulers by more than2×.

Resource Allocation: Sun et al. use a deep neural networkto approximate the mapping between the input and outputof the Weighted Minimum Mean Square Error resourceallocation algorithm [304] under interference-limited wirelessnetwork environments [290]. By effective imitation learning,the neural network approximation achieves close performanceas its teacher, while only requiring 256 shorter runtime.Deep learning has also been applied to the cloud radioaccess networks, where Xu et al. employ deep Q learning todetermine on/off mode of remote radio heads given currentmode and user demand [291]. Comparisons with Singlebase station association and fully coordinated associationmethods suggest that the proposed DRL controller allows thesystem to satisfy user demand while requiring significantly

less energy. In addition, Ferreira et al. employ deep State-Action-Reward-State-Action (SARSA) to address resourceallocation management in the cognitive communications[292]. By forecasting effects of radio parameters, theirframework avoids wasted trials of bad parameters whichreduce computational resource required.

Radio Control: In [294], the authors address the dynamicspectrum access problem in multichannel wireless networkenvironments using deep reinforcement learning. In thissetting, they incorporate LSTM into a deep Q network tomaintain and memorize the historical observations, allowingthe architecture to perform precise state estimation givenpartial observations. They distribute the training process toeach users, which enables effective training parallelizationand learning good policies for individual users. Experimentsdemonstrate that this framework achieves double the channelthroughput when compared to a benchmark method. The work[295] sheds light on the radio control and signal detectionproblems. In particular, the authors introduce a radio signalsearch environment based on Gym Reinforcement Learning.Their agent exhibits steady learning process and is able tolearn a radio signal search policy.

Other applications: Beyond the application previously dis-cussed, deep learning is also playing an important role in other


network control problems. Mao et al. develop the Pensievesystem that generates adaptive video bit rate algorithms usingdeep reinforcement learning [298]. Specifically, Pensieve em-ploys a state-of-the-art deep reinforcement learning algorithmA3C, which takes the bandwidth, bit rate and buffer size asinput, and selects the best bit rate which leads to best expectedreturn. The model is trained on an offline setting and deployedon an adaptive bit rate server, demonstrating that the systemoutperforms the best existing scheme by 12%-25% in termsof QoE. This work represents the first important step towardsimplementing DRL onto real network systems. Kim links deeplearning with the load balancing problem of IoT [301]. Hesuggests that DBN can effectively analyze network load andprocess structural configuration, thereby achieving efficientload balancing in IoT.

E. Deep Learning Driven Network Security

With the increasing popularity of wireless connectivity,protecting users, network equipment and data from maliciousattacks, unauthorized access and information leakage becomescrucial. Cybersecurity systems shelter mobile devices andusers through firewalls, anti-virus software, and an IntrusionDetection System (IDS) [305]. The firewall is an accesssecurity gateway between two networks. It allows or blocksthe uplink and downlink network traffic based on pre-definedrules. Anti-virus software detect and remove computer viruses,worms and Trojans and malwares. IDSs identify unauthorizedand malicious activities or rule violations in informationsystems. Each performs its own functions to protect networkcommunication, central servers and edge devices.

Modern cyber security systems benefit increasingly fromdeep learning [334], since it can enable the system to (i)automatically learn signatures and patterns from experienceand generalize to future intrusions (supervised learning); or(ii) identify patterns that are clearly differed from regularbehavior (unsupervised learning). This dramatically reducesthe effort of pre-defined rules for discriminating intrusions.Beyond protecting networks from attacks, deep leaning canalso play the role of “attacker”, having huge potential forstealing or cracking users’ password or information. Wesummarize these work in Table XII, which we will discussnext.

Anomaly Detection: Anomaly detection, which aims at iden-tifying network events (e.g. attack, unexpected access anduse of data) that do not conform to an expected behaviors,is becoming a key technique in IDSs. Many researchers puteffort on this areas by exploiting the outstanding unsupervisedability of AEs [306]. For example, Thing investigates featuresof attacks and threats exist in IEEE 802.11 networks [139].He employs a stacked AE to categorize network traffic into5 types (i.e. legitimate, flooding, injection and impersonationtraffic), achieving 98.67% overall accuracy. The AE is alsoexploited in [307], where Aminanto and Kim use an MLP andstacked AE for feature selection and extraction, demonstratingremarkable performance. Similarly, Feng et al. use AE todetect abnormal spectrum usage in wireless communication

[308]. The experiments suggest that the detection accuracycan significantly benefit from the depth of AEs.

Distributed attack detection is also an important issue inmobile network security. Khan et al. focus on detecting flood-ing attacks in wireless mesh networks [309]. They simulatea wireless environment with 100 nodes, and artificially injectintermediate and severe distributed flooding attacks to generatesynthetic dataset. Their deep learning based methods achieveexcellent false positive and false-negative rates. Distributedattacks are also studied in [310], where the authors focus onan IoT scenario. Another work in [311] employs MLPs todetect distributed denial of service attacks. By characterizingtypical patterns of attack incidents, their model work wellin detecting both known and unknown distributed denial ofservice intrusions.

Martin et al. propose a conditional VAE to identifyintrusion incidents in IoT [312]. In order to improve detectionperformance, their VAE infers missing features associatedwith incomplete measurements, which is common in an IoTenvironment. The true data labels are embedded into thedecoder layers to assist final classification. Evaluations on thewell-known NSL-KDD dataset [335] demonstrate that theirmodel achieves remarkable accuracy in identifying Denialof service, probing, remote to user and user to root attacks,outperforming traditional ML methods by 0.18 in terms ofF1 score.

Malware Detection: Nowadays, mobile devices are carryingconsiderable amount of private information. This informationcan be stolen and exploited by malicious apps installed onsmartphones for ill-conceived purposes [336]. Deep learningis being exploited for analyzing and detecting such threats.

Yuan et al. use both labeled and unlabeled mobile apps totrain a RBM [313]. By learning from 300 samples, their modelcan classify Android malware with remarkable accuracy, out-performing traditional ML tools by up to 19%. Their follow-up research in [314] named Droiddetector further improves thedetection accuracy by 2%. Similarly, Su et al. analyze essentialfeatures of Android apps, namely requested permission, usedpermission, sensitive application programming interface calls,action and app components [315]. They employ DBNs toextract features of malware and an SVM for classification,achieving high accuracy and only requiring 6 seconds perinference instance.

Hou et al. attack the malware detection problem from adifferent perspective. Their research points out that signature-based detection is insufficient to deal with sophisticatedAndroid malware [316]. To address this problem, theypropose the Component Traversal which can automaticallyexecute code routines to construct weighted directed graphs.By employing a Stacked AE for graph analysis, theirframework Deep4MalDroid can accurately detect Androidmalware that intentionally repackage and obfuscates tobypass signatures and hinders analysis attempts to their inneroperations. This work is followed by that of Martinelli etal., who exploit CNNs to discover the relationship betweenapp types and extracted syscall traces from real mobiledevices [317]. The CNN has also been used in [318],


TABLE XII: A summary of work on deep learning driven network security.

Application Reference Application Problem considered Learningparadigm Model

Intrusion detection

Azar et al. [306] Cyber securityapplications

Malware classification & Denial of service,probing, remote to user & user to root

Unsuper-vised &

supervisedStacked AE

Thing [139]IEEE 802.11 networkanomaly detection and

attack classification

Flooding, injection and impersonationattack

Unsuper-vised &


Aminanto andKim [307]

Wi-Fi impersonationattack detection

Flooding, injection and impersonationattack

Unsuper-vised &

supervisedMLP, AE

Feng et al. [308] Spectrum anomalydetection Additive white Gaussian noise

Unsuper-vised &

supervisedAE

Khan et al. [309] Flooding attack detectionin wireless mesh networks

Intermediate and severe distributed floodattack Supervised MLP

Diro andChilamkurti [310]

IoT distributed attackdetection

Denial of service, probing, remote to user& user to root Supervised MLP

Saied et al. [311] Distributed denial ofservice attack detection

Known and unknown distributed denial ofservice attack Supervised MLP

Martin et al.[312] IoT intrusion detection Denial of service, probing, remote to user

& user to root

Unsuper-vised &

supervised

ConditionalVAE

Malware detection

Yuan et al. [313] Android malwaredetection

Apps in contagio mobile and Google PlayStore

Unsuper-vised &

supervisedRBM

Yuan et al. [314] Android malwaredetection

Apps in contagio mobile and Google PlayStore and Genome Project

Unsuper-vised &

supervisedDBN

Su et al. [315] Android malwaredetection

Apps in Drebin, Android MalwareGenome Project and the Contagio

Community and Google Play Store

Unsuper-vised &

supervisedDBN + SVM

Hou et al. [316] Android malwaredetection

App samples from Comodo Cloud SecurityCenter

Unsuper-vised &


Martinelli [317] Android malwaredetection

Apps in Drebin, Android MalwareGenome Project and Google Play Store Supervised CNN

McLaughlin et al.[318]

Android malwaredetection

Apps in Android Malware Genome projectand Google Play Store Supervised CNN

Chen et al. [319]Malicious applicationdetection in the edge

networkPublicly-available malicious applications

Unsuper-vised &

supervisedRBM

Wang et al. [175] Malware trafficclassification Traffic extracted from 9 types of malware Supervised CNN

Botnet detection

Oulehla et al.[320] Mobile botnet detection Client-server and hybrid botnets Unknown Unknown

Torres et al. [321] Botnet detection Spam, HTTP and unknown traffic Supervised LSTMEslahi et al. [322] Mobile botnet detection HTTP botnet traffic Supervised MLPAlauthaman et al.

[323]Peer-to-peer botnet

detection Waledac and Strom Bots Supervised MLP

Privacy

Shokri andShmatikov [324]

Privacy preserving deeplearning

Avoiding sharing data in collaborativemodel training Supervised MLP, CNN

Phong et al. [325] Privacy preserving deeplearning

Addressing information leakage introducedin [324] Supervised MLP

Ossia et al. [326] Privacy-preserving mobileanalytics Offloading feature extraction from cloud Supervised CNN

Abadi et al. [327] Deep learning withdifferential privacy

Preventing exposure of private informationin training data Supervised MLP

Osia et al. [328] Privacy-preservingpersonal model training Offloading personal data from clouds

Unsuper-vised &

supervised

MLP & LatentDirichlet

Allocation [329]

Servia et al. [330] Privacy-preserving modelinference

Breaking down large models forprivacy-preserving analytics Supervised CNN

Attacker

Hitaj et al. [331] Stealing information fromcollaborative deep learning

Breaking the ordinary and differentiallyprivate collaborative deep learning

Unsuper-vised GAN

Hitaj et al. [327] Password guessing Generating passwords from leakedpassword set

Unsuper-vised GAN

Greydanus [332] Enigma learning Reconstructing functions of polyalphabeticcipher Supervised LSTM

Maghrebi [333] Breaking cryptographic Side channel attacks Supervised MLP, AE,CNN, LSTM


where the authors draw inspirations from NLP, and take thedisassembled byte-code of an app as a text for analysis. Theirexperiments demonstrate that CNN can effectively learn todetect sequences of opcodes that are indicative of malware.Chen et al. incorporate location information into the detectionframework and exploit RBM for feature extraction andclassification [319]. Their proposal improve the performanceof other ML methods.

Botnet Detection: A botnet is a network that consists ofmachines compromised by bots. These machine are usuallyunder the control of a botmaster who takes advantages of botsto harm public services and systems [337]. Detecting botnetsis challenging and now becoming a pressing task in cybersecurity.

Deep learning is playing an important role in this area. Forexample, Oulehla et al. propose to employ neural networksto extract features from mobile botnet behaviors [320]. Theydesign a parallel detection framework for detecting bothclient-server and hybrid botnets and demonstrate encouragingperformance. Torres et al. investigate the common behaviorpatterns that botnets exhibit across their life cycle usingLSTM [321]. They employ both under-sampling and over-sampling to address the class imbalance between botnet andnormal traffic in the dataset, which is common in anomalydetection problems. The similar problem is also studies in[322] and [323], where the authors use standard MLPs toperform mobile and peer-to-peer botnet detection respectively,achieving high overall accuracy.

Privacy: Preserving user privacy during training and evalu-ating a deep neural network is another important researchissue [338]. Initial research is conducted in [324], wherethe authors enable participants in training and evaluating aneural network without sharing their input data. This allows topreserve individual’s privacy while benefiting all users as theycollaboratively improve the model performance. Their frame-work is revisited and improved in [325], where another groupof researchers employ additively homomorphic encryption toaddress the information leakage problem ignored in [324]without compromising model accuracy. This significantly re-inforces the security of the system.

Osia et al. focus on privacy-preserving mobile analyticsusing deep learning. They design a client-server frameworkbased on the Siamese architecture [339], which accommodatesa feature extractor in mobile devices and correspondingly aclassifier in the cloud [326]. By offloading feature extractionsfrom the cloud, their system offers strong privacy guarantees.An innovative work in [327] implies that deep neural networkscan be trained with differential privacy. The authors introducea differentially private SGD to avoid disclosure of privateinformation of training data. Experiments on two publicly-available image recognition datasets demonstrate that theiralgorithm is able to maintain users privacy, with a manageablecost in terms of complexity, efficiency, and performance.

Servia et al. consider to train deep neural networks ondistributed devices without violating privacy constraints[330]. In specific, the authors retrain an initial model locally

tailored to individual users. This avoids transferring personaldata to untrusted entities hence users’ privacy is guaranteed.Osia et al. dedicated to protecting user’s personal data frominferences’ perspective. In particular, they break the entiredeep neural network into a feature extractor (on the client side)and an analyzer (on the cloud side) to minimize the exposureof sensitive information. Through local processing of rawinput data, sensitive personal information is transferred intoabstract features which avoid direct disclosure to the cloud.Experiments on gender classification and emotion detectionsuggest that their framework can effectively preserve users’privacy, while maintaining remarkable inference accuracy.

Attacker: Though having less applications, deep learninghas been for security attack, such as compromising user’sprivate information to guessing a enigma. In [331], Hitaj etal. suggest that collaborative learning a deep model is notreliable. By training a GAN, their attacker is able to affectthe collaboratively learning process and lure the victims todisclose private information by injecting fake training samples.Their GAN even successfully breaks the differentially privatecollaborative learning in [327]. The authors further investigatethe use of GANs for password guessing. In [340], theydesign a PassGAN to learn the distribution of a set of leakedpasswords. Once trained on a dataset, the PassGAN is ableto match over 46% of passwords in a different testing set,without user intervention or cryptography knowledge. Thisnovel technique has potential to revolutionize current passwordguessing algorithms.

Greydanus break a decryption rule using a LSTM network[332]. They treat decryption as a sequence-to-sequence trans-lation task, and train a framework with large enigma pairs.The proposed LSTM demonstrates remarkable performancein learning polyalphabetic ciphers. Maghrebi et al. exploitvarious deep learning models (i.e. MLP, AE, CNN, LSTM) toconstruct a precise profiling system to perform side channelkey recovery attacks [333]. Surprisingly, deep learning basedmethods demonstrate overwhelming performance over othertemplate machine learning attacks in terms of efficiency inbreaking both unprotected and protected Advanced EncryptionStandard implementations.

F. Emerging Deep Learning Driven Mobile Network Applica-tions

In this part, we review work based on deep learning onother mobile networking areas, which are beyond the scopesof all aforementioned subjects. These emerging applicationsare very interesting and open several new research directions.A short summary abstracts these work in Table XIII.

Deep Learning Driven Signal Processing: Deep learning isalso gaining increasing attentions in the signal processing area,especially in Multi-Input Multi-Output (MIMO) applications.The MIMO is a technique that enables cooperations andutilizations of multiple radio antennas in wireless networks.Multi-user MIMO systems aim at accommodating a largenumber of users and devices simultaneously within a


condensed area while maintaining high throughputs andconsistent performance. This is now a fundamental techniquein current wireless communications, including both cellularand WiFi networks. By incorporating deep learning, MIMOis evolving to a more intelligent system which can optimizeits performance based on aware environments.

Samuel et al. suggest that deep neural network can be agood estimator for transmitted vectors in a MIMO channel.By unfolding a projected gradient descent method, they designa MLP-based detection network to perform binary MIMOdetection [341]. The Detection Network can be implementedon multiple channels after a single training. Simulationsconducted demonstrate that their architecture achieves near-optimal accuracy while requiring light computation withoutprior knowledge of Signal-to-Noise Ratio (SNR). Yan et al.employ deep learning to solve the similar problem from adifferent perspective [342]. By considering the characteristicinvariance of signals, they exploit an AE as a feature extractor,and subsequently use an Extreme Learning Machine (ELM)to classify signal sources in a MIMO orthogonal frequencydivision multiplexing system. Their proposal achieves higherdetection accuracy over several traditional methods whileyielding similar complexity.

Wijaya et al. consider applying deep learning to a differ-ent scenario [296], [297]. The authors propose to use non-iterative neural networks to perform transmit power control atbase stations, thereby preventing the degeneration of networkperformance introduced by inter-cell interference. The neuralnetwork is trained to estimate the probability of transmitpower. At every packet transmission, the transmit power withthe highest activation probability will be selected. Simulationsdemonstrate that their framework significantly outperform thebelief propagation algorithm that routinely used for transmitpower control in MIMO environments, while yielding lesscomputational cost.

Deep learning is enabling progress in radio signal analysis.In [343], O’Shea et al. employ an LSTM to replace sequencetranslation routines between radio transmitter and receiver.Though their framework works well in ideal environments,its performance drops significantly when introducing realisticchannel effects. Later, they consider a different scenario in[344], where they exploit a regularized AE to enable reliablecommunications over over an impaired channel betweensignal senders and receivers. They further incorporate aradio transformer network on the decoder side for signalreconstruction, thereby achieving receiver synchronization.Simulations demonstrate that their framework is reliableand can be efficiently implemented. West and O’Shea laterturn their attention to modulation recognition. In [345], theycompare the recognition accuracy of different deep learningarchitectures, including traditional CNNs, ResNet, InceptionCNN and LSTM. Their experiments suggest that LSTM is thebest candidate for modulation recognition since it achievesthat highest accuracy. Due to its superior performance, LSTMis also employed for the similar task in [346].

Network Data Monetization: Gonzalez and Vallina employ

unsupervised deep learning to generate real-time accurate userprofiles [351] using an on-network machine learning platformNet2Vet [355]. Specifically, they analyze user browsingdata in real time and generate user profiles using productcategories. The profiles can be subsequently associated withthe products that are of interest to the users and then usesuch information for online advertising.

IoT In-Network Computation: Instead of taking IoT nodesas producers of data or the end-consumer of processed infor-mation, Kaminski et al. embed neural networks into a IoTnetwork, allowing IoT nodes to collaboratively process datagenerated in synchronization with the data flows [352]. Thisenables low-latency communication in IoT networks, whileoffloading data storage and processing from the cloud. Inparticular, they map each hidden unit of a per-trained neuralnetwork as a node in IoT networks, and investigate the op-timal projection which leads to the minimum communicationoverhead. Their framework performs a similar function on in-network computation in WSNs, which opens a new researchdirections in fog computing.Mobile Crowdsensing: Xiao et al. advocate that thereexist malicious mobile users who intentionally provide falsesensing data to servers for cost saving and privacy preserving,making mobile crowdsensings system vulnerable [353]. Theyformulate the server-users system as a Stackelberg game,where the server plays the role of leader that is responsiblefor evaluating the sensing effort of individuals, by analyzingthe accuracy of each sensing report. Users are paid bythe evaluations of the effort, hence cheating users will bepunished with zero reward. To design the optimal paymentpolicy, the servers employs a deep Q network to deriveknowledge from experience sensing reports and paymentpolicy without requiring knowledges of specific sensingmodels. Simulations demonstrate superior performance interms of sensing quality, resilience to attacks and server utilityover traditional Q learning and random payment strategies.

VII. TAILORING DEEP LEARNING TO MOBILE NETWORKS

Although deep learning performs remarkably in many mo-bile networking areas, the No Free Lunch (NFL) theoremindicates that there is no single model that can work univer-sally well in all problems [356]. This implies that for anyspecific mobile and wireless networking problem, we mayneed to adapt a different deep learning architectures so as toachieve better performance. In this section, we focus on how totailor deep learning to mobile network applications from threeperspectives, namely, mobile devices and systems, distributeddata centers, and changing mobile network environments.

A. Tailoring Deep Learning to Mobile Devices and Systems

The ultra-low latency requirements of future 5G mobilenetworks demand with runtime efficiency of operations inmobile systems, including deep learning driven applications.However, running complex deep learning on mobile systemsmay violate certain latency constrains. On the other hand,


TABLE XIII: A summary of emerging deep learning driven mobile network applications.

Reference Application ModelSamuel et al. [341] MIMO detection MLP

Yan et al. [342] Signal detection on a MIMO-OFDM system AE+ELMVieira et al. [263] Massive MIMO fingerprint-based positioning CNN

Neumann et al. [347] MIMO channel estimation CNNWijaya et al. [296], [297] Intercell-interference cancellation and transmit power optimization RBM

O’Shea et al. [348] Optimisation of representations and encoding/decoding processes AEO’Shea et al. [348] Sparse linear inverse problem of MIMO signal LAMP, LVAMP, CNNO’Shea et al. [343] Radio traffic sequence recognition LSTMO’Shea et al. [344] Learning to communicate over an impaired channel AE + radio transformer network

Rajendran et al. [346] Automatic modulation classification LSTMWest and O’Shea [345] Modulation recognition CNN, ResNet, Inception CNN, LSTM

O’Shea et al. [349] Modulation recognition Radio transformer networkO’Shea and Hoydis [350] Modulation classification CNN

Gonzalez and Vallina [351] Network data monetisation UnknownKaminski et al. [352] In-network computation for IoT MLP

Xiao et al. [353] Mobile crowdsensing Deep Q learningDörner et al. [354] Continuous data transmission AE

TABLE XIV: Summary of works on deep learning for mobile devices and systems.

Reference Methods Target modelIandola et al. [360] Filter size shrinking, reducing input channels and late downsampling CNNHoward et al. [361] Depth-wise separable convolution CNNZhang et al. [362] Point-wise group convolution and channel shuffle CNNZhang et al. [363] Tucker decomposition AECao et al. [364] Data parallelisation by RenderScript RNN

Chen et al. [365] Space exploration for data reusability and kernel redundancy removal CNNRallapalli et al. [366] Memory optimisations CNN

Lane et al. [367] Runtime layer compression and deep architecture decomposition MLP, CNNHuynh et al. [368] Caching, Tucker decomposition and computation offloading CNN

Wu et al. [369] Parameters quantisation CNNBhattacharya and Lane [370] Sparsification of fully-connected layers and separation of convolutional kernels MLP, CNN

Georgiev et al. [82] Representation sharing MLPCho and Brand [371] Convolution operation optimisation CNN

Guo and Potkonjak [372] Filters and classes pruning CNNLi et al. [373] Cloud assistance and incremental learning CNN

Zen et al. [374] Weight quantisation LSTMFalcao et al. [375] Parallelisation and memory sharing Stacked AE

most current mobile devices are limited by the capabilityof hardware. This means that implementing complex deeplearning architectures on such equipment without tuning maybe computationally infeasible. To address this issue, numerousresearchers dedicated to improve existing deep learning archi-tectures [357], such that they will not violate any latency andenergy constraint [358], [359]. Before the review, we outlinethese works in Table XIV.

Iandola et al. design a compacted SqueezeNet for embeddedsystems The SqueezeNet gains similar accuracy compared toAlexNet while embracing 50 times less parameters [360].Howard et al. extend this work and introduce an efficientfamily of streamlined CNNs called MobileNet, which usesdepth-wise separable convolution operations to drastically re-duce computation required and model size [361]. This newdesign results in small models which can run with lowlatency, enabling them to satisfy the requirements for mobileand embedded vision applications. They further introducetwo hyper-parameters to control the width and resolutionof multipliers which can draw appropriate trade-off betweenaccuracy and efficiency. The ShuffleNet proposed by Zhang etal. improves the accuracy of MobileNet by employing point-wise group convolution and channel shuffle while retaining

similar model complexity [362]. They discover that moregroups of convolution can reduce the computation require-ment, hence one can increase the group number to expand theinformation encoded by models for performance improvementunder certain computational capacity constrains.

Zhang et al. focus on reducing parameters of fully-connected layers for mobile multimedia features learning[363]. By applying Trucker decomposition to weight sub-tensors in the model, the dimensionality of parameters issignificantly reduced while maintaining decent reconstructioncapability. The Trucker decomposition has also been employedin [368], where the authors seek to approximate the model withless parameters for memory saving. Mobile optimizations arefurther studied for RNN models. In [364], Cao et al. use amobile toolbox called RenderScript 16 to parallelize specificdata structures and enable mobile GPUs to perform computa-tional accelerations. Their proposal significantly reduces thelatency of running RNN models on Android smartphones.Chen et al. shed light on implementing CNN on iOS mobiledevices [365]. In particular, they reduce the latency for modelexecutions, namely space exploration for data re-usability

16Android Renderscript ttps://developer.android.com/guide/topics/renderscript/compute.html.


and kernel redundancy removal. The former alleviates thehigh bandwidth requirement of convolutional layers while thelatter reduces the memory and computational requirementswith negligible performance degenerations. These drasticallyreduces overhead and latency for running CNN on mobiledevices.

Rallapalli et al. investigate offloading very deep CNNs fromclouds to edge devices by using memory optimization on bothmobile CPUs and GPUs [366]. Their framework allows to rundeep CNNs with large memory requirements at high speedon a mobile object detection application. Lane et al. developa software accelerator DeepX to assist deep learning imple-mentations on mobile devices by exploiting two inference-timeresource control algorithms, i.e. runtime layer compression anddeep architecture decomposition [367]. Specifically, runtimelayer compression technique controls the the memory andcomputation runtime during the inference phase, by extendingmodel compression principles. This is important in mobiledevices, since offloading inference to edges is more practicalon current hardware platforms. Further, the deep architecturedesigns “decomposition plans” which seeks to allocate dataand model operations to local and remote processors optimally,tailored to individual neural network structures. By combingthese two, the DeepX enables maximization of energy andruntime efficiency under certain computation and memoryconstraints.

Beyond these works, researchers also successfully adaptdeep learning architectures through other designs and sophis-ticated optimizations, such as parameters quantization [369],[374], sparsification and separation [370], representation andmemory sharing [82], [375], convolution operation optimiza-tion [371], pruning [372], and cloud assistance [373]. Thesetechniques will be of great significance to embed deep neuralnetworks into mobile systems and devices.

B. Tailoring Deep Learning to Distributed Data Containers

Mobile systems generate and consume massive mobile dataevery day, which may involve similar content but distributedaround the world. Moving all these data to centralized serversto perform model training and evaluation inevitably introducescommunication and storage overheads, which is difficult toscale. However, mobile data generated from different locationsusually exhibits different characteristics associated with humanculture, mobility, geographical topology, etc. To obtain arobust deep learning model for mobile network applications,training the model with diverse data becomes necessary. More-over, completely accommodating the full training/inferenceprocess on a cloud will introduce non-negligible computationaloverheads. Hence, appropriately offloading model executionsto distributed data centers or edge devices can dramaticallyalleviate the burden on the cloud.

Generally, there exist two solutions to address this problem.Namely, (i) decomposing the model itself to train (or makeinference with) its components individually; or (ii) scalingthe training process to perform model update at differentlocations associated with data containers. Both schemes allowone to train a single model without requiring to centralize all

data. We illustrate the principle of these two solutions in Fig.8 and review related techniques in this subsection.

Model Parallelism Large-scale distributed deep learning isfirst studied in [95], where the authors develop a frameworknamed DistBelief, which enables training complex neuralnetworks on thousands of machines. In their framework,the full model is partitioned into smaller components anddistributed over various machines. Only nodes with edges (e.g.connections between layers) that cross boundaries betweenmachines require communications for parameters update andinference. This system further involves a parameter serverwhich enables each model replica to obtain latest parametersduring training. Experiments demonstrate that the proposedframework achieves significantly faster training speed on aCPU cluster over a single GPU, while achieving state-of-the-art performance on the ImageNet [144] classification.

Teerapittayanon et al. propose distributed deep neural net-works tailored to distributed systems, which include cloudservers, fog layers and geographically distributed devices[376]. Particularly, they scale the overall neural network ar-chitecture and distribute its components hierarchically fromcloud to end devices. The model exploits local aggregatorsand binary weights to reduce computational storage, andcommunication overheads, while maintaining decent accuracy.Experiments on a multi-view multi-camera dataset demon-strate their proposal can perform efficient cloud-based trainingand local inference over a distributed computing system. Im-portantly, without violating latency constrains, the distributeddeep neural network obtains essential benefits associated withdistributed systems, such as fault tolerance and privacy.

Coninck et al. consider distributing deep learning overIoT for classification applications [377]. Specifically, theydeploy a small neural network to local devices to performcoarse classification, which enables fast respond filtered datato be sent to central servers. If the local model fails toclassify, the larger neural network on the cloud is activatedto perform fine-grained classification. The overall architecturemaintain comparable accuracy, while significantly reducinglatency introduced by large model inference.

Decentralized methods can also be applied to deepreinforcement learning. In [378], Omidshafiei et al. considera multi-agent system with partial observability and limitedallowed communication, which is common in mobile networksystems. They combine a set of sophisticated methods andalgorithms, including hysteretic learners, deep recurrentQ network, concurrent experience replay trajectories anddistillation, to enable multi-agent coordination using a singlejoint policy under a set of decentralized partial observableMDPs. Their framework can potentially play an importantrole in addressing control problems in distributed mobilesystems. As controllers in such systems have strict limitedobservability and strict communication constraints, the wholeprocess can be formulated as a partially observable MDP.

Training Parallelism Training parallelism is also essential formobile system, as mobile data usually come asynchronouslyfrom different sources. Effectively training models while main-


Machine/Device 1

Machine/Device 2

Machine/Device 3 Machine/Device 4

Model Parallelism

Data Collection

Asynchronous SGD

Time

Training Parallelism

Fig. 8: The underlying principles of model parallelism (left) and training parallelism (right).

taining consistency, fast convergence, and accuracy remainschallenging [379].

A practical method to address this problem is to performasynchronous SGD. The basic idea is to enable the serverthat maintains a model to allow to accept stale delayedinformation (e.g. data, gradient update) from workers. At eachupdate iteration, the server only requires to wait for a smallernumber of workers. This is essential for training a deep neuralnetwork over distributed machines in mobile systems. Theasynchronous SGD is first studied in [380], where the authorspropose a lock-free parallel SGD named HOGWILD, whichdemonstrates significant faster convergence over locking coun-terparts. The Downpour SGD in [95] improves the robustnessof the training process when work nodes breakdown, aseach model replica requests the latest version of parameters.Hence small number of failed machines will not make asignificant impact to the training process. A similar idea hasbeen employed in [381], where Goyal et al. investigate theusage of a set of techniques (i.e. learning rate adjustment,warm-up, batch normalization) of large minibatches whichoffer important insights on training large-scale deep neuralnetworks on distributed system. Eventually, their frameworkcan train an accurate network on ImageNet in 1 hour, whichis impressive in comparison with traditional algorithms.

Zhang et al. advocate that most of asynchronous SGD algo-rithms suffer from slow convergence due to the inherent vari-ance of stochastic gradients [382]. They propose an improvedSGD with variance reduction to speed up the convergence.Their algorithm significantly outperforms other asynchronousSGD in terms of convergence when training deep neuralnetworks on Google Cloud Computing Platform. The asyn-chronous method has also been applied to deep reinforcementlearning. In [66], the authors create multiple environmentswhich allows agents to perform asynchronous updates to themain structure. The new algorithm A3C breaks the sequentialdependency, reduces the reliability of experience replay, andsignificantly speeds up the training of the traditional Actor-

Critic algorithm. In [383], Hardy et al. further study learninga neural network in a distributed manner over cloud andedge devices. In particular, they propose a training algorithm,AdaComp, which allows to compress worker updates to thetargeted model. This significantly reduce the communicationoverhead between cloud and edge, while retaining good faulttolerance with the occurrence of worker breakdowns.

Federated learning has recently become an emerging paral-lelism approach that enables mobile devices to collaborativelylearn a shared model while retaining all training data onindividual device [384], [385]. Beyond offloading trainingdata from central servers, it performs model update underSecure Aggregation protocol [386], which decrypts the averageupdate only enough users have participated without inspectindividual phone’s update. This allows model to aggregateupdates aggregation in a secure, efficient, scalable, and fault-tolerant way.

C. Tailoring Deep Learning to Changing Mobile NetworkEnvironments

Mobile network environments usually exhibit changingpatterns over time. For instance, spatial mobile data trafficdistributions over a region significantly vary from differenttime of a day [387]. Applying a deep learning model inchanging mobile environments will requires lifelong learningability to continuously absorb new features from mobileenvironments, without forgetting old but essential patterns.Moreover, new smartphone-targeted viruses are spreading fastvia mobile network and severely jeopardize users’ privacyand commercial profits. These pose unprecedented challengesto current anomaly detection systems and anti-virus software,as they are required such frameworks to timely react to newthreats using limited information. To this end, the modelshould have transfer learning ability, which can enable to fasttransfer the knowledge from pre-trained models for differentjobs or dataset. This will allow models to work well with


Learning Task 1 Learning task 2 Learning Task n

...

Time

Deep Learning Model

Knowledge

Base

Knowledge 1 Knowledge 2 Knowledge n

...

SupportSupport

Time

Knowledge 1 Knowledge 2 Knowledge nKnowledge

TransferKnowledge

Transfer

Model 1 Model 2 Model n

Deep Lifelong Learning Deep Transfer Learning

...

Learning Task 1 Learning task 2 Learning Task n

...

Fig. 9: The underlying principles of deep lifelong learning (left) and deep transfer learning (right). The lifelong learning retainthe knowledge learned while transfer learning exploits the source domain labeled data to help target domain learning withoutknowledge retention.

limited threat samples (one-shot learning) or limited metadatadescription of new threats (zero-shot learning). Therefore,both lifelong learning and transfer learning are essential forapplications in changeable mobile network environments. Weillustrated these two learning paradigms in Fig. 9 and reviewessential research in this subsection.

Deep Lifelong Learning Lifelong learning mimics humanbehaviors and seeks to build a machine that can continuouslyadapt to new environments [388]. An ideal lifelong learningmachine is able to retain knowledge from previous learningexperience to aid future learning and problem solving withnecessary adjustments, which is of great significance in learn-ing problems in mobile network environments.

There exist several research efforts which adapt traditionaldeep learning to a lifelong learning machine. For example,Lee et al. propose a dual-memory deep learning architec-ture for lifelong learning of everyday human behaviors overnon-stationary data streams [389]. To enable the pre-trainedmodel to retain old knowledge while training with new data,their architecture includes two memory buffers, namely deepmemory and fast memory. The deep memory is composed ofseveral deep networks, which are built when the amount ofdata from an unseen distribution is accumulated and reachesa threshold. The fast memory component is a small neuralnetwork, which is updated immediately when coming acrossa new data sample. These two memory modules allow toperform continuous learning without forgetting old knowledge.Experiments on a non-stationary image data stream provethe effectiveness of this model, as it significantly outper-forms other online deep learning algorithms. The memorymechanism has also been applied in [390]. In particular,the authors introduce a differentiable neural computer, whichallows neural network to dynamically read from and write toan external memory module. This enables lifelong lookup andforgetting of knowledge from external sources, as humans do.

Parisi et al. consider a different lifelong learning scenario

in [391]. In particular, they abandon the memory modulesin [389] and design a self-organizing architecture with withrecurrent neurons for processing time-varying patterns. Avariant of the Growing When Required network is employedin each layer to to predict neural activation sequences from theprevious network layer, which allows learning the time-varycorrelations between input and label, without requirements ofa predefined number of classes. Importantly, their frameworkis robust, as it has tolerance to missing and corrupted samplelabels which is common in mobile data.

Another interesting deep lifelong learning architecture ispresented in [392], where Tessler et al. manage to build aDQN agent which can retain learned skills in playing thefamous computer game Minecraft. The overall frameworkincludes a pre-trained model, Deep Skill Network, whichis trained a-priori on various sub-tasks of the game. Whenlearning a new task, the old knowledge is maintained byincorporating reusable skills through a Deep Skill module,which consists of a Deep Skill Network array and a multi-skilldistillation network. These allow the agent to selectivelytransfer knowledge to solve a new task. Experimentsdemonstrate that their proposal significantly outperformstraditional double DQN in terms of accuracy and convergenceby reusing and transferring old skills in new task learning.This technique has potential to be employed in solving mobilenetworking problems, as it can continuously new knowledge.

Deep Transfer Learning Unlike lifelong learning, the transferlearning only seeks to use knowledge from a specific domainto aid target domain learning. Applying transfer learning canaccelerate the new learning process, as the new task does notrequire to learn from scratch. This is essential to mobile net-work environments, as they require to agilely respond to newnetwork patterns and threats. Now it has important applicationin computer network domain [51], such as Web mining [393],caching [394] and base station sleeping mechanism [158].

There exist two extreme transfer learning paradigms,


namely one-shot learning and zero-shot learning. The one-shot learning refers to a learning method that gains as muchinformation as possible about a category from only one or ahandful of samples, given a pre-trained model [395]. On theother hand, zero-shot learning does not require any samplefrom a category [396]. It aims at learning a new distributiongiven meta description of the new category and correlationwith existing training data. Though research toward deepone-shot learning [80], [397] and deep zero-shot learning[398], [399] are novel, both paradigms are quite promisingin detecting new threats or patterns in mobile networks.

VIII. FUTURE RESEARCH PERSPECTIVES

Although deep learning is achieving increasingly promisingresults in the mobile networking domain, several essentialopen research issues exist and are worthy of attention inthe future. In what follows, we discuss these challenges andpinpoint important mobile networking problems that can besolved by deep learning. These can deliver insights into futuremobile networking research.

A. Serving Deep Learning with Massive and High-QualityData

Deep neural networks rely on massive and high-qualitydata to gain satisfying performance. When training a largeand complex architecture, data volume and quality are veryimportant, as deeper model usually has a huge set of param-eters to be learned and configured. This issue remains truein mobile network applications. Unfortunately, unlike somepopular research areas such as computer vision and NLP,there still lacks high-quality and large-scale labeled datasetsfor mobile network applications, as service provides preferto keep their data confidential and reluctant to release theirdatasets. This to some extent, restricts the development ofdeep learning in the mobile networking domain. Moreover,due to limitations of mobile sensors and network equipment,mobile data collected are usually subjected to loss, redundancy,mislabeling and class imbalance, and thus cannot be directlyemployed for training purpose.

To establish an intelligent 5G mobile network architecture,efficient and mature streamlining and platforms for mobile dataprocessing are in demand, to serve deep learning applications.This requires considerable amount of research efforts for datacollection, transmission, cleaning, clustering and transforma-tion. We appeal to researchers and companies to release moredatasets, which can dramatically advance the deep learningapplications in the mobile network area and benefit a widerange of communities from both academic and industry.

B. Deep Learning in Spatio-Temporal Mobile Traffic DataMining

Accurate analysis of mobile traffic data over a geographicalregion is becoming increasingly essential for event localiza-tion, network resource allocation, context-based advertisingand urban planning [387]. However, due to the mobility ofsmartphone users, the spatio-temporal distribution of mobile

traffic and application popularity is highly irregular (see anexample in Fig. 10) [400], [401], which is particularly difficultto capture their complex correlations.

3D Mobile Traffic Surface 2D Mobile Traffic Snapshot

Fig. 10: An example of 3D mobile traffic surface (left) and2D projection (right) on Milan, Italy. Figures adapted from[163] using data from [402].

Recent research suggests that data collected by mobilesensors (e.g. mobile traffic) over a city can be regardedas pictures taken by panoramic cameras, which provide acity-scale sensing system for urban surveillance [403]. Thesetraffic sensing images enclose information associated withmovements of large-scale of individuals [162]. We recognizethat mobile traffic data, from both spatial and temporal di-mensions, have important similarity with videos, images orspeech sequence, as been confirmed in [163]. We illustratetheir analogies in Fig. 11.

t

...

t + s

Mobile Traffic EvolutionsVideo

Image

Speech Signal

Mobile Traffic

Snapshot

Mobile Traffic

Series

Zooming

Time

Tra

ffic

volu

me

Frequency

Amplitute

Longitude

Latitu

de

LongitudeLongitude

Latitu

de

Latitu

de

Fig. 11: Analogies between mobile traffic data (left) and otherdata (right).

Specifically, the evolution of large -scale mobile traffichighly resemble videos, as they are both composed of asequence of “frames”. If we focus on individual traffic snap-shot, the spatial mobile traffic distribution is by analogy withimages. Moreover, if we zoom into a narrow location to mea-sure its long-term traffic consumption, we can observe that asingle traffic series looks similar to natural language sequence.These imply that, to some extent, well-established tools for


computer vision (e.g. CNN) or NLP (e.g. RNN, LSTM) canbe a promising candidate for mobile traffic analysis [163].

Beyond the similarity, we observe several properties thatmobile traffic snapshots uniquely embrace, making them dis-tinct from images or language sequences. Namely,

1) Values of neighboring pixels in fine-grained traffic snap-shots normally do not change significantly, while thishappens quite often in edge areas of natural images.

2) Single mobile traffic series usually exhibits regular peri-odicity (in both daily and weekly properties), but suchproperty does not hold for pixels in videos.

3) Due to users’ mobility, the mobile traffic consumption ina region is more likely to stay or shift to neighboringareas in the near future, while the values of pixels mightnot be shifted in videos.

Such properties of mobile traffic simplify their spatio-temporalcorrelations, which can be exploited as prior knowledge formodel design. We recognize several unique advantages ofemploying deep learning for mobile traffic mining:

1) Deep learning performs remarkably in imaging and NLPapplications, while mobile traffic has significant analogiesto these data but with less spatio-temporal complexity;

2) LSTM captures well temporal correlations and depen-dency of time series data, while irregular and stochasticuser mobility could be learned using the transformationabilities of deformable CNN [146].

3) Advanced GPU computing enables fast training of neuralnetworks and parallelism techniques can support low-latency mobile traffic analysis.

We expect that deep learning can effectively exploited thespatio-temporal dynamic to solve limitation of existing toolssuch as Exponential Smoothing [404] and AutoregressiveIntegrated Moving Average model [405]. This may raise theresearch on mobile traffic traffic analysis to a higher level.

C. Deep Unsupervised Learning Powered Mobile Data Anal-ysis

We observe that current deep learning practices in mo-bile networks largely employ supervised learning and rein-forcement learning, while the potential of deep unsupervisedlearning is yet to be explored. However, as mobile networksgenerate considerable amounts of unlabeled data every day,data labeling is costly and requires domain-specific knowl-edge. To facilitate the analysis of raw mobile network data,unsupervised learning becomes essential in extracting insightsfrom unlabeled data [406], so as to optimize the mobilenetwork functionality to improve QoE.

Deep learning is resourceful in terms of unsupervisedlearning, as it provides excellent unsupervised tools such asAE, RBM and GAN. These models in general require lightfeature engineering thus being promising in learning fromheterogeneous and unstructured mobile data. In particular,deep AEs work well unsupervised anomaly detection. One canuse normal data only to train an AE, and take test data thatyields significantly higher reconstruction error than trainingerror as outliers [407]. Though less popular, RBM can performlayer-wise unsupervised pre-training which can accelerate the

overall model training process. GANs are good at imitatingdata distributions, which can be employed as a simulator tomimic real mobile network environments. Recent researchreveals that GANs can protect communications by craftingcustom their own cryptography to avoid eavesdropping [408].All these tools require further research to fulfill their fullpotentials in the mobile networking domain.

D. Deep Reinforcement Learning Powered Mobile NetworkControl

Currently, many mobile network control problems have beensolved by constrained optimization, dynamic programmingand game theory approaches. Unfortunately, these methodseither make strong assumption about the objective functions(e.g. function convexity) or data distribution (e.g. Gaussianor Poisson distributed), or suffer from high time and spacecomplexity. However, as mobile networks become increasinglycomplex, such assumptions sometimes turn unrealistic. Theobjective functions may be further is affected by large setof variables which poses severe computational and memorychallenges to these mathematical approaches.

On the other hand, despite the Markov property, deepreinforcement learning does not make strong assumptions tothe target system. It employs function approximation whichperfectly addresses the problem of large state-action space,enabling reinforcement learning to scale to network controlproblems that were previously intractable. Inspired by itsremarkable achievements in Atari [128] and the game of Go[409], a few researchers start to bring DRL to solve complexnetwork control problems, which has been reviewed in Sec.VI-D. However, we believe that these work only displays asmall part of DRL’s talent and its potential in mobile networkcontrol remains largely unexplored. For instance, it can beexploited to extract rich features from cellular networks toenable intelligent switching on/off base stations to reduceinfrastructure’s energy footprint [410]. This should be, asDeepMind trains a DRL agent which reduces Google datacenter cooling bill by 40% 17. These exciting applicationsmake us believe DRL will raise the performance of mobilenetwork control to a higher level in the future.

IX. CONCLUSIONS

Deep learning is playing an increasingly important role inthe mobile and wireless networking domain. In this paper,we provided a comprehensive survey of recent work thatlies at the intersection between these two different areas. Wesummarized both basic concepts and advanced principles ofvarious deep learning models, then correlated the deep learningand mobile networking disciplines by reviewing work acrossdifferent application scenarios. We discussed how to tailordeep learning models to general mobile networking appli-cations, an aspect entirely overlooked by previous surveys.We concluded by pinpointing several open research issuesand promising directions, which may lead to valuable future

17DeepMind AI Reduces Google Data Center Cooling Bill by 40%https://deepmind.com/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-40/


research results. We hope this article will become a definiteguide to researchers and practitioners interested in applyingapplying machine intelligence to to complex problems inmobile network environments.

ACKNOWLEDGEMENT

We would like to thank Zongzuo Wang for sharing valuableinsights on deep learning, which helped improving the qualityof this paper.

REFERENCES

[1] Cisco. Cisco Visual Networking Index: Forecast and Methodology,2016-2021, June 2017.

[2] Ning Wang, Ekram Hossain, and Vijay K Bhargava. Backhauling 5Gsmall cells: A radio resource management perspective. IEEE WirelessCommunications, 22(5):41–49, 2015.

[3] Fabio Giust, Luca Cominardi, and Carlos J Bernardos. Distributedmobility management for future 5G networks: overview and analysisof existing approaches. IEEE Communications Magazine, 53(1):142–149, 2015.

[4] Mamta Agiwal, Abhishek Roy, and Navrati Saxena. Next generation5G wireless networks: A comprehensive survey. IEEE CommunicationsSurveys & Tutorials, 18(3):1617–1655, 2016.

[5] Akhil Gupta and Rakesh Kumar Jha. A survey of 5G network:Architecture and emerging technologies. IEEE access, 3:1206–1232,2015.

[6] Kan Zheng, Zhe Yang, Kuan Zhang, Periklis Chatzimisios, Kan Yang,and Wei Xiang. Big data-driven optimization for mobile networkstoward 5G. IEEE network, 30(1):44–51, 2016.

[7] Chunxiao Jiang, Haijun Zhang, Yong Ren, Zhu Han, Kwang-ChengChen, and Lajos Hanzo. Machine learning paradigms for next-generation wireless networks. IEEE Wireless Communications,24(2):98–105, 2017.

[8] Duong D Nguyen, Hung X Nguyen, and Langford B White. Rein-forcement learning with network-assisted feedback for heterogeneousrat selection. IEEE Transactions on Wireless Communications, 2017.

[9] Fairuz Amalina Narudin, Ali Feizollah, Nor Badrul Anuar, and Ab-dullah Gani. Evaluation of machine learning classifiers for mobilemalware detection. Soft Computing, 20(1):343–357, 2016.

[10] Kevin Hsieh, Aaron Harlap, Nandita Vijaykumar, Dimitris Konomis,Gregory R Ganger, Phillip B Gibbons, and Onur Mutlu. Gaia: Geo-distributed machine learning approaching LAN speeds. In USENIXSymposium on Networked Systems Design and Implementation (NSDI),pages 629–647, 2017.

[11] Wencong Xiao, Jilong Xue, Youshan Miao, Zhen Li, Cheng Chen,Ming Wu, Wei Li, and Lidong Zhou. Tux2: Distributed graph com-putation for machine learning. In USENIX Symposium on NetworkedSystems Design and Implementation (NSDI), pages 669–682, 2017.

[12] Paolini, Monica and Fili, Senza . Mastering Analytics: How to benefitfrom big data and network complexity: An Analyst Report. RCRWireless News, 2017.

[13] Anastasia Ioannidou, Elisavet Chatzilari, Spiros Nikolopoulos, andIoannis Kompatsiaris. Deep learning advances in computer vision with3d data: A survey. ACM Computing Surveys (CSUR), 50(2):20, 2017.

[14] Richard Socher, Yoshua Bengio, and Christopher D Manning. Deeplearning for nlp (without magic). In Tutorial Abstracts of ACL 2012,pages 5–5. Association for Computational Linguistics.

[15] IEEE Network special issue: Exploring Deep Learning for Efficientand Reliable Mobile Sensing. http://www.comsoc.org/netmag/cfp/exploring-deep-learning-efficient-and-reliable-mobile-sensing, 2017.[Online; accessed 14-July-2017].

[16] Mowei Wang, Yong Cui, Xin Wang, Shihan Xiao, and Junchen Jiang.Machine learning for networking: Workflow, advances and opportuni-ties. IEEE Network, 2017.

[17] Mohammad Abu Alsheikh, Dusit Niyato, Shaowei Lin, Hwee-PinkTan, and Zhu Han. Mobile big data analytics using deep learningand Apache Spark. IEEE network, 30(3):22–29, 2016.

[18] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning.Nature, 521(7553):436–444, 2015.

[19] Jürgen Schmidhuber. Deep learning in neural networks: An overview.Neural networks, 61:85–117, 2015.

[20] Weibo Liu, Zidong Wang, Xiaohui Liu, Nianyin Zeng, Yurong Liu,and Fuad E Alsaadi. A survey of deep neural network architecturesand their applications. Neurocomputing, 234:11–26, 2017.

[21] Li Deng, Dong Yu, et al. Deep learning: methods and applications.Foundations and Trends R© in Signal Processing, 7(3–4):197–387,2014.

[22] Li Deng. A tutorial survey of architectures, algorithms, and applicationsfor deep learning. APSIPA Transactions on Signal and InformationProcessing, 3, 2014.

[23] Ian Goodfellow, Yoshua Bengio, and Aaron Courville. Deep learning.MIT press, 2016.

[24] Kai Arulkumaran, Marc Peter Deisenroth, Miles Brundage, andAnil Anthony Bharath. A brief survey of deep reinforcement learn-ing. arXiv:1708.05866, 2017. To appear in IEEE Signal ProcessingMagazine, Special Issue on Deep Learning for Image Understanding.

[25] Xue-Wen Chen and Xiaotong Lin. Big data deep learning: challengesand perspectives. IEEE access, 2:514–525, 2014.

[26] Maryam M Najafabadi, Flavio Villanustre, Taghi M Khoshgoftaar,Naeem Seliya, Randall Wald, and Edin Muharemagic. Deep learningapplications and challenges in big data analytics. Journal of Big Data,2(1):1, 2015.

[27] NF Hordri, A Samar, SS Yuhaniz, and SM Shamsuddin. A systematicliterature review on features of deep learning in big data analytics. In-ternational Journal of Advances in Soft Computing & Its Applications,9(1), 2017.

[28] Mehdi Gheisari, Guojun Wang, and Md Zakirul Alam Bhuiyan. Asurvey on deep learning in big data. In Computational Science andEngineering (CSE) and Embedded and Ubiquitous Computing (EUC),IEEE International Conference on, volume 2, pages 173–180, 2017.

[29] Shui Yu, Meng Liu, Wanchun Dou, Xiting Liu, and Sanming Zhou.Networking for big data: A survey. IEEE Communications Surveys &Tutorials, 19(1):531–549, 2017.

[30] Mohammad Abu Alsheikh, Shaowei Lin, Dusit Niyato, and Hwee-PinkTan. Machine learning in wireless sensor networks: Algorithms, strate-gies, and applications. IEEE Communications Surveys & Tutorials,16(4):1996–2018, 2014.

[31] Chun-Wei Tsai, Chin-Feng Lai, Ming-Chao Chiang, Laurence T Yang,et al. Data mining for Internet of things: A survey. IEEE Communi-cations Surveys and Tutorials, 16(1):77–97, 2014.

[32] Xiang Cheng, Luoyang Fang, Xuemin Hong, and Liuqing Yang.Exploiting mobile big data: Sources, features, and applications. IEEENetwork, 31(1):72–79, 2017.

[33] Mario Bkassiny, Yang Li, and Sudharman K Jayaweera. A survey onmachine-learning techniques in cognitive radios. IEEE Communica-tions Surveys & Tutorials, 15(3):1136–1159, 2013.

[34] Jeffrey G Andrews, Stefano Buzzi, Wan Choi, Stephen V Hanly, AngelLozano, Anthony CK Soong, and Jianzhong Charlie Zhang. Whatwill 5G be? IEEE Journal on selected areas in communications,32(6):1065–1082, 2014.

[35] Nisha Panwar, Shantanu Sharma, and Awadhesh Kumar Singh. Asurvey on 5G: The next generation of mobile communication. PhysicalCommunication, 18:64–84, 2016.

[36] Olakunle Elijah, Chee Yen Leow, Tharek Abdul Rahman, SolomonNunoo, and Solomon Zakwoi Iliya. A comprehensive survey of pilotcontamination in massive MIMO–5G system. IEEE CommunicationsSurveys & Tutorials, 18(2):905–923, 2016.

[37] Stefano Buzzi, I Chih-Lin, Thierry E Klein, H Vincent Poor, ChenyangYang, and Alessio Zappone. A survey of energy-efficient techniquesfor 5G networks and challenges ahead. IEEE Journal on Selected Areasin Communications, 34(4):697–709, 2016.

[38] Mugen Peng, Yong Li, Zhongyuan Zhao, and Chonggang Wang.System architecture and key technologies for 5G heterogeneous cloudradio access networks. IEEE network, 29(2):6–14, 2015.

[39] Yong Niu, Yong Li, Depeng Jin, Li Su, and Athanasios V Vasilakos.A survey of millimeter wave communications (mmwave) for 5G:opportunities and challenges. Wireless Networks, 21(8):2657–2676,2015.

[40] Xenofon Foukas, Georgios Patounas, Ahmed Elmokashfi, and Ma-hesh K Marina. Network slicing in 5G: Survey and challenges. IEEECommunications Magazine, 55(5):94–100, 2017.

[41] Tarik Taleb, Konstantinos Samdanis, Badr Mada, Hannu Flinck, SunnyDutta, and Dario Sabella. On multi-access edge computing: A surveyof the emerging 5G network edge architecture & orchestration. IEEECommunications Surveys & Tutorials, 2017.

[42] Pavel Mach and Zdenek Becvar. Mobile edge computing: A surveyon architecture and computation offloading. IEEE CommunicationsSurveys & Tutorials, 2017.


[43] Yuyi Mao, Changsheng You, Jun Zhang, Kaibin Huang, and Khaled BLetaief. A survey on mobile edge computing: The communicationperspective. IEEE Communications Surveys & Tutorials, 2017.

[44] Ying Wang, Peilong Li, Lei Jiao, Zhou Su, Nan Cheng, Xuemin Sher-man Shen, and Ping Zhang. A data-driven architecture for personalizedQoE management in 5G wireless networks. IEEE Wireless Communi-cations, 24(1):102–110, 2017.

[45] Qilong Han, Shuang Liang, and Hongli Zhang. Mobile cloud sensing,big data, and 5G networks make an intelligent and smart world. IEEENetwork, 29(2):40–45, 2015.

[46] Sukhdeep Singh, Navrati Saxena, Abhishek Roy, and HanSeok Kim.f. IETE Technical Review, 34(1):30–39, 2017.

[47] Min Chen, Jun Yang, Yixue Hao, Shiwen Mao, and Kai Hwang. A 5Gcognitive system for healthcare. Big Data and Cognitive Computing,1(1):2, 2017.

[48] Teodora Sandra Buda, Haytham Assem, Lei Xu, Danny Raz, UdiMargolin, Elisha Rosensweig, Diego R Lopez, Marius-Iulian Corici,Mikhail Smirnov, Robert Mullins, et al. Can machine learning aid indelivering new use cases and scenarios in 5G? In Network Operationsand Management Symposium (NOMS), 2016 IEEE/IFIP, pages 1279–1284, 2016.

[49] Ali Imran, Ahmed Zoha, and Adnan Abu-Dayya. Challenges in 5G:how to empower SON with big data for enabling 5G. IEEE Network,28(6):27–33, 2014.

[50] Bharath Keshavamurthy and Mohammad Ashraf. Conceptual designof proactive SONs based on the big data framework for 5G cellularnetworks: A novel machine learning perspective facilitating a shift inthe son paradigm. In System Modeling & Advancement in ResearchTrends (SMART), International Conference, pages 298–304. IEEE,2016.

[51] Paulo Valente Klaine, Muhammad Ali Imran, Oluwakayode Onireti,and Richard Demo Souza. A survey of machine learning techniquesapplied to self organizing cellular networks. IEEE CommunicationsSurveys and Tutorials, 2017.

[52] Rongpeng Li, Zhifeng Zhao, Xuan Zhou, Guoru Ding, Yan Chen,Zhongyao Wang, and Honggang Zhang. Intelligent 5G: When cellularnetworks meet artificial intelligence. IEEE Wireless Communications,2017.

[53] Nicola Bui, Matteo Cesana, S Amir Hosseini, Qi Liao, Ilaria Malan-chini, and Joerg Widmer. A survey of anticipatory mobile networking:Context-based classification, prediction methodologies, and optimiza-tion techniques. IEEE Communications Surveys & Tutorials, 2017.

[54] Panagiotis Kasnesis, Charalampos Patrikakis, and Iakovos Venieris.Changing the game of mobile data analysis with deep learning. ITProfessional, 2017.

[55] Xiang Cheng, Luoyang Fang, Liuqing Yang, and Shuguang Cui. Mobilebig data: The fuel for data-driven wireless. IEEE Internet of ThingsJournal, 2017.

[56] Lidong Wang and Randy Jones. Big data analytics for networkintrusion detection: A survey. International Journal of Networks andCommunications, 7(1):24–31, 2017.

[57] Nei Kato, Zubair Md Fadlullah, Bomin Mao, Fengxiao Tang, OsamuAkashi, Takeru Inoue, and Kimihiro Mizutani. The deep learning visionfor heterogeneous network traffic control: proposal, challenges, andfuture perspective. IEEE Wireless Communications, 24(3):146–153,2017.

[58] Michele Zorzi, Andrea Zanella, Alberto Testolin, Michele De FilippoDe Grazia, and Marco Zorzi. Cognition-based networks: A newperspective on network optimization using learning and distributedintelligence. IEEE Access, 3:1512–1530, 2015.

[59] Zubair Fadlullah, Fengxiao Tang, Bomin Mao, Nei Kato, OsamuAkashi, Takeru Inoue, and Kimihiro Mizutani. State-of-the-art deeplearning: Evolving machine intelligence toward tomorrow’s intelligentnetwork traffic control systems. IEEE Communications Surveys &Tutorials, 2017.

[60] Mehdi Mohammadi, Ala Al-Fuqaha, Sameh Sorour, and MohsenGuizani. Deep learning for IoT big data and streaming analytics: Asurvey. arXiv preprint arXiv:1712.04301, 2017.

[61] Nauman Ahad, Junaid Qadir, and Nasir Ahsan. Neural networks inwireless networks: Techniques, applications and guidelines. Journal ofNetwork and Computer Applications, 68:1–27, 2016.

[62] Ammar Gharaibeh, Mohammad A Salahuddin, Sayed J Hussini, Ab-dallah Khreishah, Issa Khalil, Mohsen Guizani, and Ala Al-Fuqaha.Smart cities: A survey on data management, security and enablingtechnologies. IEEE Communications Surveys & Tutorials, 2017.

[63] Nicholas D Lane and Petko Georgiev. Can deep learning revolutionizemobile sensing? In Proceedings of the 16th International Workshop on

Mobile Computing Systems and Applications, pages 117–122. ACM,2015.

[64] Kaoru Ota, Minh Son Dao, Vasileios Mezaris, and Francesco GBDe Natale. Deep learning for mobile multimedia: A survey. ACMTransactions on Multimedia Computing, Communications, and Appli-cations (TOMM), 13(3s):34, 2017.

[65] Shuai Zhang, Lina Yao, and Aixin Sun. Deep learning based rec-ommender system: A survey and new perspectives. arXiv preprintarXiv:1707.07435, 2017.

[66] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, AlexGraves, Timothy Lillicrap, Tim Harley, David Silver, and KorayKavukcuoglu. Asynchronous methods for deep reinforcement learning.In International Conference on Machine Learning, pages 1928–1937,2016.

[67] Martin Arjovsky, Soumith Chintala, and Leon Bottou. Wasserstein gen-erative adversarial networks. In International Conference on MachineLearning (2017), 2017.

[68] W. McCulloch and W. Pitts. A logical calculus of the ideas immanentin nervous activity. Bulletin of Mathematical Biophysics, (5).

[69] DRGHR Williams and Geoffrey Hinton. Learning representations byback-propagating errors. Nature, 323(6088):533–538, 1986.

[70] Yann LeCun, Yoshua Bengio, et al. Convolutional networks for images,speech, and time series. The handbook of brain theory and neuralnetworks, 3361(10):1995, 1995.

[71] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. Imagenetclassification with deep convolutional neural networks. In Advances inneural information processing systems, pages 1097–1105, 2012.

[72] Pedro Domingos. A few useful things to know about machine learning.Commun. ACM, 55(10):78–87, 2012.

[73] Ivor W Tsang, James T Kwok, and Pak-Ming Cheung. Core vectormachines: Fast SVM training on very large data sets. Journal ofMachine Learning Research, 6:363–392, 2005.

[74] Carl Edward Rasmussen and Christopher KI Williams. Gaussianprocesses for machine learning, volume 1. MIT press Cambridge,2006.

[75] Nicolas Le Roux and Yoshua Bengio. Representational power ofrestricted boltzmann machines and deep belief networks. Neuralcomputation, 20(6):1631–1649, 2008.

[76] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, DavidWarde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Gen-erative adversarial nets. In Advances in neural information processingsystems, pages 2672–2680, 2014.

[77] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: Aunified embedding for face recognition and clustering. In Proceedingsof the IEEE Conference on Computer Vision and Pattern Recognition,pages 815–823, 2015.

[78] Diederik P Kingma, Shakir Mohamed, Danilo Jimenez Rezende, andMax Welling. Semi-supervised learning with deep generative models.In Advances in Neural Information Processing Systems, pages 3581–3589, 2014.

[79] Russell Stewart and Stefano Ermon. Label-free supervision of neuralnetworks with physics and domain knowledge. In AAAI, pages 2576–2582, 2017.

[80] Danilo Rezende, Ivo Danihelka, Karol Gregor, Daan Wierstra, et al.One-shot generalization in deep generative models. In InternationalConference on Machine Learning, pages 1521–1529, 2016.

[81] Richard Socher, Milind Ganjoo, Christopher D Manning, and AndrewNg. Zero-shot learning through cross-modal transfer. In Advances inneural information processing systems, pages 935–943, 2013.

[82] Petko Georgiev, Sourav Bhattacharya, Nicholas D Lane, and CeciliaMascolo. Low-resource multi-task audio sensing for mobile andembedded devices via shared deep neural network representations. Pro-ceedings of the ACM on Interactive, Mobile, Wearable and UbiquitousTechnologies, 1(3):50, 2017.

[83] Norman P Jouppi, Cliff Young, Nishant Patil, David Patterson, GauravAgrawal, Raminder Bajwa, Sarah Bates, Suresh Bhatia, Nan Boden,Al Borchers, et al. In-datacenter performance analysis of a tensorprocessing unit. arXiv preprint arXiv:1704.04760, 2017.

[84] John Nickolls, Ian Buck, Michael Garland, and Kevin Skadron. Scal-able parallel programming with CUDA. Queue, 6(2):40–53, 2008.

[85] Sharan Chetlur, Cliff Woolley, Philippe Vandermersch, Jonathan Co-hen, John Tran, Bryan Catanzaro, and Evan Shelhamer. cuDNN:Efficient primitives for deep learning. arXiv preprint arXiv:1410.0759,2014.

[86] Martín Abadi, Paul Barham, Jianmin Chen, Zhifeng Chen, Andy Davis,Jeffrey Dean, Matthieu Devin, Sanjay Ghemawat, Geoffrey Irving,


Michael Isard, et al. TensorFlow: A system for large-scale machinelearning. In OSDI, volume 16, pages 265–283, 2016.

[87] Theano Development Team. Theano: A Python framework forfast computation of mathematical expressions. arXiv e-prints,abs/1605.02688, May 2016.

[88] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev,Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell.Caffe: Convolutional architecture for fast feature embedding. arXivpreprint arXiv:1408.5093, 2014.

[89] R. Collobert, K. Kavukcuoglu, and C. Farabet. Torch7: A Matlab-likeenvironment for machine learning. In BigLearn, NIPS Workshop, 2011.

[90] Vinayak Gokhale, Jonghoon Jin, Aysegul Dundar, Berin Martini, andEugenio Culurciello. A 240 G-ops/s mobile coprocessor for deep neuralnetworks. In Proceedings of the IEEE Conference on Computer Visionand Pattern Recognition Workshops, pages 682–687, 2014.

[91] ncnn is a high-performance neural network inference framework opti-mized for the mobile platform . https://github.com/Tencent/ncnn, 2017.[Online; accessed 25-July-2017].

[92] Huawei announces the Kirin 970- new flagship SoC with AI capabil-ities. http://www.androidauthority.com/huawei-announces-kirin-970-797788/, 2017. [Online; accessed 01-Sep-2017].

[93] Core ML: Integrate machine learning models into your app. https://developer.apple.com/documentation/coreml, 2017. [Online; accessed25-July-2017].

[94] Ilya Sutskever, James Martens, George E Dahl, and Geoffrey E Hinton.On the importance of initialization and momentum in deep learning.ICML (3), 28:1139–1147, 2013.

[95] Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin,Mark Mao, Andrew Senior, Paul Tucker, Ke Yang, Quoc V Le,et al. Large scale distributed deep networks. In Advances in neuralinformation processing systems, pages 1223–1231, 2012.

[96] Diederik Kingma and Jimmy Ba. Adam: A method for stochasticoptimization. International Conference on Learning Representations(ICLR), 2015.

[97] Tim Kraska, Ameet Talwalkar, John C Duchi, Rean Griffith, Michael JFranklin, and Michael I Jordan. MLbase: A distributed machine-learning system. In CIDR, volume 1, pages 2–1, 2013.

[98] Trishul M Chilimbi, Yutaka Suzue, Johnson Apacible, and KarthikKalyanaraman. Project adam: Building an efficient and scalable deeplearning training system. In OSDI, volume 14, pages 571–582, 2014.

[99] Henggang Cui, Hao Zhang, Gregory R Ganger, Phillip B Gibbons, andEric P Xing. Geeps: Scalable deep learning on distributed GPUs witha GPU-specialized parameter server. In Proceedings of the EleventhEuropean Conference on Computer Systems, page 4. ACM, 2016.

[100] Ryan Spring and Anshumali Shrivastava. Scalable and sustainabledeep learning via randomized hashing. ACM SIGKDD Conference onKnowledge Discovery and Data Mining, 2017.

[101] Azalia Mirhoseini, Hieu Pham, Quoc V Le, Benoit Steiner, RasmusLarsen, Yuefeng Zhou, Naveen Kumar, Mohammad Norouzi, SamyBengio, and Jeff Dean. Device placement optimization with reinforce-ment learning. International Conference on Machine Learning, 2017.

[102] Eric P Xing, Qirong Ho, Wei Dai, Jin Kyu Kim, Jinliang Wei, SeunghakLee, Xun Zheng, Pengtao Xie, Abhimanu Kumar, and Yaoliang Yu.Petuum: A new platform for distributed machine learning on big data.IEEE Transactions on Big Data, 1(2):49–67, 2015.

[103] Moustafa Alzantot, Yingnan Wang, Zhengshuang Ren, and Mani BSrivastava. RSTensorFlow: GPU enabled tensorflow for deep learningon commodity Android devices. In Proceedings of the 1st InternationalWorkshop on Deep Learning for Mobile Systems and Applications,pages 7–12. ACM, 2017.

[104] Hao Dong, Akara Supratak, Luo Mai, Fangde Liu, Axel Oehmichen,Simiao Yu, and Yike Guo. TensorLayer: A versatile library for efficientdeep learning development. In Proceedings of the 2017 ACM onMultimedia Conference, MM ’17, pages 1201–1204, 2017.

[105] Adam Paszke, Sam Gross, Soumith Chintala, Gregory Chanan, EdwardYang, Zachary DeVito, Zeming Lin, Alban Desmaison, Luca Antiga,and Adam Lerer. Automatic differentiation in pytorch. 2017.

[106] Tianqi Chen, Mu Li, Yutian Li, Min Lin, Naiyan Wang, Minjie Wang,Tianjun Xiao, Bing Xu, Chiyuan Zhang, and Zheng Zhang. Mxnet:A flexible and efficient machine learning library for heterogeneousdistributed systems. arXiv preprint arXiv:1512.01274, 2015.

[107] Sebastian Ruder. An overview of gradient descent optimizationalgorithms. arXiv preprint arXiv:1609.04747, 2016.

[108] Marcin Andrychowicz, Misha Denil, Sergio Gomez, Matthew W Hoff-man, David Pfau, Tom Schaul, and Nando de Freitas. Learning tolearn by gradient descent by gradient descent. In Advances in NeuralInformation Processing Systems, pages 3981–3989, 2016.

[109] Wei Wen, Cong Xu, Feng Yan, Chunpeng Wu, Yandan Wang, YiranChen, and Hai Li. TernGrad: Ternary gradients to reduce communi-cation in distributed deep learning. In Advances in neural informationprocessing systems, 2017.

[110] Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed,Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and AndrewRabinovich. Going deeper with convolutions. In Proceedings of theIEEE conference on computer vision and pattern recognition, pages1–9, 2015.

[111] Flavio Bonomi, Rodolfo Milito, Preethi Natarajan, and Jiang Zhu. Fogcomputing: A platform for internet of things and analytics. In BigData and Internet of Things: A Roadmap for Smart Environments,pages 169–186. Springer, 2014.

[112] Jiachen Mao, Xiang Chen, Kent W Nixon, Christopher Krieger, andYiran Chen. MoDNN: Local distributed mobile computing system fordeep neural network. In 2017 Design, Automation & Test in EuropeConference & Exhibition (DATE), pages 1396–1401. IEEE, 2017.

[113] Suyoung Bang, Jingcheng Wang, Ziyun Li, Cao Gao, Yejoong Kim,Qing Dong, Yen-Po Chen, Laura Fick, Xun Sun, Ron Dreslinski, et al.14.7 a 288µw programmable deep-learning processor with 270kb on-chip weight storage using non-uniform memory hierarchy for mobileintelligence. In IEEE International Conference on Solid-State Circuits(ISSCC), pages 250–251, 2017.

[114] Filipp Akopyan. Design and tool flow of ibm’s truenorth: an ultra-low power programmable neurosynaptic chip with 1 million neurons.In Proceedings of the 2016 on International Symposium on PhysicalDesign, pages 59–60. ACM, 2016.

[115] Seyyed Salar Latifi Oskouei, Hossein Golestani, Matin Hashemi, andSoheil Ghiasi. Cnndroid: GPU-accelerated execution of trained deepconvolutional neural networks on Android. In Proceedings of the 2016ACM on Multimedia Conference, pages 1201–1205. ACM, 2016.

[116] Corinna Cortes, Xavi Gonzalvo, Vitaly Kuznetsov, Mehryar Mohri, andScott Yang. Adanet: Adaptive structural learning of artificial neuralnetworks. ICML, 2017.

[117] Geoffrey E Hinton, Simon Osindero, and Yee-Whye Teh. A fast learn-ing algorithm for deep belief nets. Neural computation, 18(7):1527–1554, 2006.

[118] Honglak Lee, Roger Grosse, Rajesh Ranganath, and Andrew Y Ng.Convolutional deep belief networks for scalable unsupervised learningof hierarchical representations. In Proceedings of the 26th annualinternational conference on machine learning, pages 609–616. ACM,2009.

[119] Pascal Vincent, Hugo Larochelle, Isabelle Lajoie, Yoshua Bengio, andPierre-Antoine Manzagol. Stacked denoising autoencoders: Learninguseful representations in a deep network with a local denoising cri-terion. Journal of Machine Learning Research, 11(Dec):3371–3408,2010.

[120] Diederik P Kingma and Max Welling. Auto-encoding variational bayes.International Conference on Learning Representations (ICLR), 2014.

[121] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deepresidual learning for image recognition. In Proceedings of the IEEEconference on computer vision and pattern recognition, pages 770–778,2016.

[122] Shuiwang Ji, Wei Xu, Ming Yang, and Kai Yu. 3D convolutional neuralnetworks for human action recognition. IEEE transactions on patternanalysis and machine intelligence, 35(1):221–231, 2013.

[123] Gao Huang, Zhuang Liu, Kilian Q Weinberger, and Laurens van derMaaten. Densely connected convolutional networks. IEEE Conferenceon Computer Vision and Pattern Recognition, 2017.

[124] Felix A Gers, Jürgen Schmidhuber, and Fred Cummins. Learning toforget: Continual prediction with LSTM. 1999.

[125] Ilya Sutskever, Oriol Vinyals, and Quoc V Le. Sequence to sequencelearning with neural networks. In Advances in neural informationprocessing systems, pages 3104–3112, 2014.

[126] Shi Xingjian, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-KinWong, and Wang-chun Woo. Convolutional LSTM network: A machinelearning approach for precipitation nowcasting. In Advances in neuralinformation processing systems, pages 802–810, 2015.

[127] Guo-Jun Qi. Loss-sensitive generative adversarial networks on Lips-chitz densities. arXiv preprint arXiv:1701.06264, 2017.

[128] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu,Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller,Andreas K Fidjeland, Georg Ostrovski, et al. Human-level controlthrough deep reinforcement learning. Nature, 518(7540):529–533,2015.

[129] David Silver, Aja Huang, Chris J Maddison, Arthur Guez, Lau-rent Sifre, George Van Den Driessche, Julian Schrittwieser, Ioannis


Antonoglou, Veda Panneershelvam, Marc Lanctot, et al. Masteringthe game of Go with deep neural networks and tree search. Nature,529(7587):484–489, 2016.

[130] Ronan Collobert and Samy Bengio. Links between perceptrons, MLPsand SVMs. In Proceedings of the twenty-first international conferenceon Machine learning, page 23. ACM, 2004.

[131] Geoffrey E Hinton. Training products of experts by minimizingcontrastive divergence. Neural computation, 14(8):1771–1800, 2002.

[132] George Casella and Edward I George. Explaining the gibbs sampler.The American Statistician, 46(3):167–174, 1992.

[133] Takashi Kuremoto, Masanao Obayashi, Kunikazu Kobayashi, TakaomiHirata, and Shingo Mabu. Forecast chaotic time series data byDBNs. In Image and Signal Processing (CISP), 2014 7th InternationalCongress on, pages 1130–1135. IEEE, 2014.

[134] Yann Dauphin and Yoshua Bengio. Stochastic ratio matching of RBMsfor sparse high-dimensional inputs. In Advances in Neural InformationProcessing Systems, pages 1340–1348, 2013.

[135] Tara N Sainath, Brian Kingsbury, Bhuvana Ramabhadran, Petr Fousek,Petr Novak, and Abdel-rahman Mohamed. Making deep belief net-works effective for large vocabulary continuous speech recognition. InAutomatic Speech Recognition and Understanding (ASRU), 2011 IEEEWorkshop on, pages 30–35, 2011.

[136] Yoshua Bengio et al. Learning deep architectures for AI. Foundationsand trends R© in Machine Learning, 2(1):1–127, 2009.

[137] Mayu Sakurada and Takehisa Yairi. Anomaly detection using au-toencoders with nonlinear dimensionality reduction. In ProceedingsWorkshop on Machine Learning for Sensory Data Analysis (MLSDA),page 4. ACM, 2014.

[138] Miguel Nicolau, James McDermott, et al. A hybrid autoencoder anddensity estimation model for anomaly detection. In InternationalConference on Parallel Problem Solving from Nature, pages 717–726.Springer, 2016.

[139] Vrizlynn LL Thing. IEEE 802.11 network anomaly detection andattack classification: A deep learning approach. In IEEE WirelessCommunications and Networking Conference (WCNC), pages 1–6,2017.

[140] Bomin Mao, Zubair Md Fadlullah, Fengxiao Tang, Nei Kato, OsamuAkashi, Takeru Inoue, and Kimihiro Mizutani. Routing or computing?the paradigm shift towards intelligent computer network packet trans-mission based on deep learning. IEEE Transactions on Computers,2017.

[141] Valentin Radu, Nicholas D Lane, Sourav Bhattacharya, Cecilia Mas-colo, Mahesh K Marina, and Fahim Kawsar. Towards multimodal deeplearning for activity recognition on mobile devices. In Proceedingsof ACM International Joint Conference on Pervasive and UbiquitousComputing: Adjunct, pages 185–188, 2016.

[142] Ramachandra Raghavendra and Christoph Busch. Learning deeply cou-pled autoencoders for smartphone based robust periocular verification.In Image Processing (ICIP), 2016 IEEE International Conference on,pages 325–329, 2016.

[143] Jing Li, Jingyuan Wang, and Zhang Xiong. Wavelet-based stackeddenoising autoencoders for cell phone base station user number predic-tion. In IEEE International Conference on Internet of Things (iThings)and IEEE Green Computing and Communications (GreenCom) andIEEE Cyber, Physical and Social Computing (CPSCom) and IEEESmart Data (SmartData), pages 833–838, 2016.

[144] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, SanjeevSatheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla,Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet LargeScale Visual Recognition Challenge. International Journal of ComputerVision (IJCV), 115(3):211–252, 2015.

[145] Junmo Kim Yunho Jeon. Active convolution: Learning the shapeof convolution for image classification. In Proceedings of the IEEEConference on Computer Vision and Pattern Recognition, 2017.

[146] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu,and Yichen Wei. Deformable convolutional networks. arXiv preprintarXiv:1703.06211, 2017.

[147] Yoshua Bengio, Patrice Simard, and Paolo Frasconi. Learning long-term dependencies with gradient descent is difficult. IEEE transactionson neural networks, 5(2):157–166, 1994.

[148] Alex Graves, Navdeep Jaitly, and Abdel-rahman Mohamed. Hybridspeech recognition with deep bidirectional LSTM. In Automatic SpeechRecognition and Understanding (ASRU), 2013 IEEE Workshop on,pages 273–278, 2013.

[149] Rie Johnson and Tong Zhang. Supervised and semi-supervised textcategorization using LSTM for region embeddings. In InternationalConference on Machine Learning, pages 526–534, 2016.

[150] Francisco Javier Ordóñez and Daniel Roggen. Deep convolutionaland LSTM recurrent neural networks for multimodal wearable activityrecognition. Sensors, 16(1):115, 2016.

[151] Christian Ledig, Lucas Theis, Ferenc Huszár, Jose Caballero, AndrewCunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Jo-hannes Totz, Zehan Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. In IEEE Conferenceon Computer Vision and Pattern Recognition, 2017.

[152] Jianan Li, Xiaodan Liang, Yunchao Wei, Tingfa Xu, Jiashi Feng, andShuicheng Yan. Perceptual generative adversarial networks for smallobject detection. In IEEE Conference on Computer Vision and PatternRecognition, 2017.

[153] Yijun Li, Sifei Liu, Jimei Yang, and Ming-Hsuan Yang. Generativeface completion. In IEEE Conference on Computer Vision and PatternRecognition, 2017.

[154] Shixiang Gu, Timothy Lillicrap, Ilya Sutskever, and Sergey Levine.Continuous deep Q-learning with model-based acceleration. In Inter-national Conference on Machine Learning, pages 2829–2838, 2016.

[155] Matej Moravcík, Martin Schmid, Neil Burch, Viliam Lisy, DustinMorrill, Nolan Bard, Trevor Davis, Kevin Waugh, Michael Johanson,and Michael Bowling. Deepstack: Expert-level artificial intelligence inheads-up no-limit poker. Science, 356(6337):508–513, 2017.

[156] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and DeirdreQuillen. Learning hand-eye coordination for robotic grasping with deeplearning and large-scale data collection. The International Journal ofRobotics Research, page 0278364917710318, 2016.

[157] Ahmad EL Sallab, Mohammed Abdou, Etienne Perot, and SenthilYogamani. Deep reinforcement learning framework for autonomousdriving. Electronic Imaging, 2017(19):70–76, 2017.

[158] Rongpeng Li, Zhifeng Zhao, Xianfu Chen, Jacques Palicot, and Hong-gang Zhang. Tact: A transfer actor-critic learning framework for energysaving in cellular radio access networks. IEEE Transactions on WirelessCommunications, 13(4):2000–2011, 2014.

[159] Hasan AA Al-Rawi, Ming Ann Ng, and Kok-Lim Alvin Yau. Ap-plication of reinforcement learning to routing in distributed wirelessnetworks: a review. Artificial Intelligence Review, 43(3):381–416, 2015.

[160] Yan-Jun Liu, Li Tang, Shaocheng Tong, CL Philip Chen, and Dong-Juan Li. Reinforcement learning design-based adaptive tracking controlwith less learning parameters for nonlinear discrete-time MIMO sys-tems. IEEE Transactions on Neural Networks and Learning Systems,26(1):165–176, 2015.

[161] Demetrios Zeinalipour Yazti and Shonali Krishnaswamy. Mobile bigdata analytics: research, practice, and opportunities. In Mobile DataManagement (MDM), 2014 IEEE 15th International Conference on,volume 1, pages 1–2. IEEE, 2014.

[162] Diala Naboulsi, Marco Fiore, Stephane Ribot, and Razvan Stanica.Large-scale mobile traffic analysis: a survey. IEEE CommunicationsSurveys & Tutorials, 18(1):124–161, 2016.

[163] Chaoyun Zhang, Xi Ouyang, and Paul Patras. ZipNet-GAN: Inferringfine-grained mobile traffic patterns via a generative adversarial neuralnetwork. In Proceedings of the 13th ACM Conference on EmergingNetworking Experiments and Technologies. ACM, 2017.

[164] Victor C Liang, Richard TB Ma, Wee Siong Ng, Li Wang, MarianneWinslett, Huayu Wu, Shanshan Ying, and Zhenjie Zhang. Mercury:Metro density prediction with recurrent neural network on streamingCDR data. In Data Engineering (ICDE), 2016 IEEE 32nd InternationalConference on, pages 1374–1377, 2016.

[165] Jiquan Ngiam, Aditya Khosla, Mingyu Kim, Juhan Nam, Honglak Lee,and Andrew Y Ng. Multimodal deep learning. In Proceedings of the28th international conference on machine learning (ICML-11), pages689–696, 2011.

[166] Laura Pierucci and Davide Micheli. A neural network for quality ofexperience estimation in mobile communications. IEEE MultiMedia,23(4):42–49, 2016.

[167] Youngjune Gwon and HT Kung. Inferring origin flow patterns in Wi-Fiwith deep learning. In ICAC, pages 73–83, 2014.

[168] Laisen Nie, Dingde Jiang, Shui Yu, and Houbing Song. Network trafficprediction based on deep belief network in wireless mesh backbonenetworks. In Wireless Communications and Networking Conference(WCNC), 2017 IEEE, pages 1–5, 2017.

[169] Vusumuzi Moyo et al. The generalization ability of artificial neuralnetworks in forecasting TCP/IP traffic trends: How much does the sizeof learning rate matter? International Journal of Computer Scienceand Application, 2015.

[170] Jing Wang, Jian Tang, Zhiyuan Xu, Yanzhi Wang, Guoliang Xue, XingZhang, and Dejun Yang. Spatiotemporal modeling and prediction incellular networks: A big data enabled deep learning approach. In


INFOCOM–36th Annual IEEE International Conference on ComputerCommunications, 2017.

[171] Chaoyun Zhang and Paul Patras. Long-term mobile traffic fore-casting using deep spatio-temporal neural networks. arXiv preprintarXiv:1712.08083, 2017.

[172] Zhanyi Wang. The applications of deep learning on traffic identifica-tion. BlackHat USA, 2015.

[173] Wei Wang, Ming Zhu, Jinlin Wang, Xuewen Zeng, and ZhongzhenYang. End-to-end encrypted traffic classification with one-dimensionalconvolution neural networks. In Intelligence and Security Informatics(ISI), 2017 IEEE International Conference on, pages 43–48, 2017.

[174] Mohammad Lotfollahi, Ramin Shirali, Mahdi Jafari Siavoshani, andMohammdsadegh Saberian. Deep packet: A novel approach forencrypted traffic classification using deep learning. arXiv preprintarXiv:1709.02656, 2017.

[175] Wei Wang, Ming Zhu, Xuewen Zeng, Xiaozhou Ye, and Yiqiang Sheng.Malware traffic classification using convolutional neural network forrepresentation learning. In Information Networking (ICOIN), 2017International Conference on, pages 712–717. IEEE, 2017.

[176] Bjarke Felbo, Pål Sundsøy, Alex’Sandy’ Pentland, Sune Lehmann,and Yves-Alexandre de Montjoye. Using deep learning to predictdemographics from mobile phone metadata. International Conferenceon Learning Representations (ICLR), workshop track, 2016.

[177] Nai Chun Chen, Wanqin Xie, Roy E Welsch, Kent Larson, and JennyXie. Comprehensive predictions of tourists’ next visit location based oncall detail records using machine learning and deep learning methods.In Big Data (BigData Congress), 2017 IEEE International Congresson, pages 1–6, 2017.

[178] Ziheng Lin, Mogeng Yin, Sidney Feygin, Madeleine Sheehan, Jean-Francois Paiement, and Alexei Pozdnoukhov. Deep generative modelsof urban mobility. 2017.

[179] Chang Xu, Kuiyu Chang, Khee-Chin Chua, Meishan Hu, and Zhenx-iang Gao. Large-scale Wi-Fi hotspot classification via deep learning.In Proceedings of the 26th International Conference on World WideWeb Companion, pages 857–858. International World Wide Web Con-ferences Steering Committee, 2017.

[180] Ala Al-Fuqaha, Mohsen Guizani, Mehdi Mohammadi, MohammedAledhari, and Moussa Ayyash. Internet of things: A survey on enablingtechnologies, protocols, and applications. IEEE Communications Sur-veys & Tutorials, 17(4):2347–2376, 2015.

[181] Suranga Seneviratne, Yining Hu, Tham Nguyen, Guohao Lan, SaraKhalifa, Kanchana Thilakarathna, Mahbub Hassan, and Aruna Senevi-ratne. A survey of wearable devices and challenges. IEEE Communi-cations Surveys & Tutorials, 2017.

[182] He Li, Kaoru Ota, and Mianxiong Dong. Learning IoT in edge:Deep learning for the Internet of Things with edge computing. IEEENetwork, 32(1):96–101, 2018.

[183] Nicholas D Lane, Sourav Bhattacharya, Petko Georgiev, Claudio For-livesi, and Fahim Kawsar. An early resource characterization of deeplearning on wearables, smartphones and internet-of-things devices. InProceedings of the 2015 International Workshop on Internet of Thingstowards Applications, pages 7–12. ACM, 2015.

[184] Sicong Liu and Junzhao Du. Poster: Mobiear-building an environment-independent acoustic sensing platform for the deaf using deep learning.In Proceedings of the 14th Annual International Conference on MobileSystems, Applications, and Services Companion, pages 50–50. ACM,2016.

[185] Liu Sicong, Zhou Zimu, Du Junzhao, Shangguan Longfei, Jun Han, andXin Wang. Ubiear: Bringing location-independent sound awareness tothe hard-of-hearing people with smartphones. Proceedings of the ACMon Interactive, Mobile, Wearable and Ubiquitous Technologies, 1(2):17,2017.

[186] Vasu Jindal. Integrating mobile and cloud for PPG signal selection tomonitor heart rate during intensive physical exercise. In Proceedingsof the International Workshop on Mobile Software Engineering andSystems, pages 36–37. ACM, 2016.

[187] Edward Kim, Miguel Corte-Real, and Zubair Baloch. A deep semanticmobile application for thyroid cytopathology. In Proc. SPIE, volume9789, page 97890A, 2016.

[188] Aarti Sathyanarayana, Shafiq Joty, Luis Fernandez-Luque, Ferda Ofli,Jaideep Srivastava, Ahmed Elmagarmid, Teresa Arora, and ShahradTaheri. Sleep quality prediction from wearable data using deeplearning. JMIR mHealth and uHealth, 4(4), 2016.

[189] Honggui Li and Maria Trocan. Personal health indicators by deeplearning of smart phone sensor data. In Cybernetics (CYBCONF),2017 3rd IEEE International Conference on, pages 1–5, 2017.

[190] Mohammad-Parsa Hosseini, Tuyen X Tran, Dario Pompili, Kost Elise-vich, and Hamid Soltanian-Zadeh. Deep learning with edge computingfor localization of epileptogenicity using multimodal rs-fMRI and EEGbig data. In Autonomic Computing (ICAC), 2017 IEEE InternationalConference on, pages 83–92, 2017.

[191] Cosmin Stamate, George D Magoulas, Stefan Küppers, EffrosyniNomikou, Ioannis Daskalopoulos, Marco U Luchini, Theano Mous-souri, and George Roussos. Deep learning parkinson’s from smartphonedata. In Pervasive Computing and Communications (PerCom), 2017IEEE International Conference on, pages 31–40, 2017.

[192] Tom Quisel, Luca Foschini, Alessio Signorini, and David C Kale. Col-lecting and analyzing millions of mhealth data streams. In Proceedingsof the 23rd ACM SIGKDD International Conference on KnowledgeDiscovery and Data Mining, pages 1971–1980. ACM, 2017.

[193] Usman Mahmood Khan, Zain Kabir, Syed Ali Hassan, and Syed Has-san Ahmed. A deep learning framework using passive WiFi sensingfor respiration monitoring. arXiv preprint arXiv:1704.05708, 2017.

[194] Dawei Li, Theodoros Salonidis, Nirmit V Desai, and Mooi ChooChuah. Deepcham: Collaborative edge-mediated adaptive deep learningfor mobile object recognition. In Edge Computing (SEC), IEEE/ACMSymposium on, pages 64–76, 2016.

[195] Luis Tobías, Aurélien Ducournau, François Rousseau, GrégoireMercier, and Ronan Fablet. Convolutional neural networks for objectrecognition on mobile devices: A case study. In Pattern Recognition(ICPR), 2016 23rd International Conference on, pages 3530–3535.IEEE, 2016.

[196] Parisa Pouladzadeh and Shervin Shirmohammadi. Mobile multi-foodrecognition using deep learning. ACM Transactions on MultimediaComputing, Communications, and Applications (TOMM), 13(3s):36,2017.

[197] Ryosuke Tanno, Koichi Okamoto, and Keiji Yanai. DeepFoodCam: ADCNN-based real-time mobile food recognition system. In Proceedingsof the 2nd International Workshop on Multimedia Assisted DietaryManagement, pages 89–89. ACM, 2016.

[198] Pallavi Kuhad, Abdulsalam Yassine, and Shervin Shimohammadi.Using distance estimation and deep learning to simplify calibration infood calorie measurement. In Computational Intelligence and VirtualEnvironments for Measurement Systems and Applications (CIVEMSA),2015 IEEE International Conference on, pages 1–6, 2015.

[199] Teng Teng and Xubo Yang. Facial expressions recognition based onconvolutional neural networks for mobile virtual reality. In Proceedingsof the 15th ACM SIGGRAPH Conference on Virtual-Reality Continuumand Its Applications in Industry-Volume 1, pages 475–478. ACM, 2016.

[200] Wu Liu, Huadong Ma, Heng Qi, Dong Zhao, and Zhineng Chen. Deeplearning hashing for mobile visual search. EURASIP Journal on Imageand Video Processing, 2017(1):17, 2017.

[201] Jinmeng Rao, Yanjun Qiao, Fu Ren, Junxing Wang, and Qingyun Du.A mobile outdoor augmented reality method combining deep learningobject detection and spatial relationships for geovisualization. Sensors,17(9):1951, 2017.

[202] Kazuya Ohara, Takuya Maekawa, and Yasuyuki Matsushita. Detectingstate changes of indoor everyday objects using Wi-Fi channel stateinformation. Proceedings of the ACM on Interactive, Mobile, Wearableand Ubiquitous Technologies, 1(3):88, 2017.

[203] Ming Zeng, Le T Nguyen, Bo Yu, Ole J Mengshoel, Jiang Zhu, PangWu, and Joy Zhang. Convolutional neural networks for human activityrecognition using mobile sensors. In Mobile Computing, Applicationsand Services (MobiCASE), 2014 6th International Conference on,pages 197–205. IEEE, 2014.

[204] Bandar Almaslukh, Jalal AlMuhtadi, and Abdelmonim Artoli. An ef-fective deep autoencoder approach for online smartphone-based humanactivity recognition. International Journal of Computer Science andNetwork Security (IJCSNS), 17(4):160, 2017.

[205] Xinyu Li, Yanyi Zhang, Ivan Marsic, Aleksandra Sarcevic, and Ran-dall S Burd. Deep learning for RFID-based activity recognition.In Proceedings of the 14th ACM Conference on Embedded NetworkSensor Systems CD-ROM, pages 164–175. ACM, 2016.

[206] Sourav Bhattacharya and Nicholas D Lane. From smart to deep: Robustactivity recognition on smartwatches using deep learning. In PervasiveComputing and Communication Workshops (PerCom Workshops), 2016IEEE International Conference on, pages 1–6, 2016.

[207] Antreas Antoniou and Plamen Angelov. A general purpose intelligentsurveillance system for mobile devices using deep learning. In NeuralNetworks (IJCNN), 2016 International Joint Conference on, pages2879–2886. IEEE, 2016.

[208] Saiwen Wang, Jie Song, Jaime Lien, Ivan Poupyrev, and OtmarHilliges. Interacting with Soli: Exploring fine-grained dynamic gesture


recognition in the radio-frequency spectrum. In Proceedings of the29th Annual Symposium on User Interface Software and Technology,pages 851–860. ACM, 2016.

[209] Yang Gao, Ning Zhang, Honghao Wang, Xiang Ding, Xu Ye, GuanlingChen, and Yu Cao. ihear food: Eating detection using commoditybluetooth headsets. In Connected Health: Applications, Systems andEngineering Technologies (CHASE), 2016 IEEE First InternationalConference on, pages 163–172, 2016.

[210] Jindan Zhu, Amit Pande, Prasant Mohapatra, and Jay J Han. Using deeplearning for energy expenditure estimation with wearable sensors. InE-health Networking, Application & Services (HealthCom), 2015 17thInternational Conference on, pages 501–506. IEEE, 2015.

[211] Pål Sundsøy, Johannes Bjelland, B Reme, A Iqbal, and Eaman Jahani.Deep learning applied to mobile phone data for individual incomeclassification. ICAITA doi, 10, 2016.

[212] Yuqing Chen and Yang Xue. A deep learning approach to humanactivity recognition based on single accelerometer. In Systems, Man,and Cybernetics (SMC), 2015 IEEE International Conference on, pages1488–1492, 2015.

[213] Sojeong Ha and Seungjin Choi. Convolutional neural networks forhuman activity recognition using multiple accelerometer and gyroscopesensors. In Neural Networks (IJCNN), 2016 International JointConference on, pages 381–388. IEEE, 2016.

[214] Marcus Edel and Enrico Köppe. Binarized-BLSTM-RNN based humanactivity recognition. In Indoor Positioning and Indoor Navigation(IPIN), 2016 International Conference on, pages 1–7. IEEE, 2016.

[215] Tsuyoshi OKITA and Sozo INOUE. Recognition of multiple overlap-ping activities using compositional cnn-lstm model. In Proceedingsof the 2017 ACM International Joint Conference on Pervasive andUbiquitous Computing and Proceedings of the 2017 ACM InternationalSymposium on Wearable Computers, pages 165–168. ACM, 2017.

[216] Gaurav Mittal, Kaushal B Yagnik, Mohit Garg, and Narayanan CKrishnan. Spotgarbage: smartphone app to detect garbage usingdeep learning. In Proceedings of the 2016 ACM International JointConference on Pervasive and Ubiquitous Computing, pages 940–945.ACM, 2016.

[217] Lorenzo Seidenari, Claudio Baecchi, Tiberio Uricchio, Andrea Ferra-cani, Marco Bertini, and Alberto Del Bimbo. Deep artwork detectionand retrieval for automatic context-aware audio guides. ACM Trans-actions on Multimedia Computing, Communications, and Applications(TOMM), 13(3s):35, 2017.

[218] Xiao Zeng, Kai Cao, and Mi Zhang. Mobiledeeppill: A small-footprintmobile deep learning system for recognizing unconstrained pill images.In Proceedings of the 15th Annual International Conference on MobileSystems, Applications, and Services, pages 56–67. ACM, 2017.

[219] Shuochao Yao, Shaohan Hu, Yiran Zhao, Aston Zhang, and TarekAbdelzaher. Deepsense: A unified deep learning framework fortime-series mobile sensing data processing. In Proceedings of the26th International Conference on World Wide Web, pages 351–360.International World Wide Web Conferences Steering Committee, 2017.

[220] Xiao Zeng. Mobile sensing through deep learning. In Proceedings ofthe 2017 Workshop on MobiSys 2017 Ph. D. Forum, pages 5–6. ACM,2017.

[221] Kleomenis Katevas, Ilias Leontiadis, Martin Pielot, and Joan Serrà.Practical processing of mobile sensor data for continual deep learningpredictions. In Proceedings of the 1st International Workshop on DeepLearning for Mobile Systems and Applications. ACM, 2017.

[222] Xuyu Wang, Lingjun Gao, and Shiwen Mao. PhaseFi: Phase finger-printing for indoor localization with a deep learning approach. InGlobal Communications Conference (GLOBECOM), 2015 IEEE, pages1–6, 2015.

[223] Xuyu Wang, Lingjun Gao, and Shiwen Mao. CSI phase fingerprintingfor indoor localization with a deep learning approach. IEEE Internetof Things Journal, 3(6):1113–1123, 2016.

[224] Bokai Cao, Lei Zheng, Chenwei Zhang, Philip S Yu, Andrea Piscitello,John Zulueta, Olu Ajilore, Kelly Ryan, and Alex D Leow. Deepmood:Modeling mobile phone typing dynamics for mood detection. 2017.

[225] Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, MohammadNorouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao,Klaus Macherey, et al. Google’s neural machine translation system:Bridging the gap between human and machine translation. arXivpreprint arXiv:1609.08144, 2016.

[226] Ian McGraw, Rohit Prabhavalkar, Raziel Alvarez, Montse GonzalezArenas, Kanishka Rao, David Rybach, Ouais Alsharif, Hasim Sak,Alexander Gruenstein, Françoise Beaufays, et al. Personalized speechrecognition on mobile devices. In Acoustics, Speech and Signal

Processing (ICASSP), 2016 IEEE International Conference on, pages5955–5959, 2016.

[227] Rohit Prabhavalkar, Ouais Alsharif, Antoine Bruguier, and Lan Mc-Graw. On the compression of recurrent neural networks with an appli-cation to LVCSR acoustic modeling for embedded speech recognition.In Acoustics, Speech and Signal Processing (ICASSP), 2016 IEEEInternational Conference on, pages 5970–5974, 2016.

[228] Takuya Yoshioka, Nobutaka Ito, Marc Delcroix, Atsunori Ogawa,Keisuke Kinoshita, Masakiyo Fujimoto, Chengzhu Yu, Wojciech JFabian, Miquel Espi, Takuya Higuchi, et al. The NTT CHiME-3system: Advances in speech enhancement and recognition for mobilemulti-microphone devices. In IEEE Workshop on Automatic SpeechRecognition and Understanding (ASRU), pages 436–443, 2015.

[229] Sherry Ruan, Jacob O Wobbrock, Kenny Liou, Andrew Ng, and JamesLanday. Speech is 3x faster than typing for english and mandarin textentry on mobile devices. arXiv preprint arXiv:1608.07323, 2016.

[230] Andrey Ignatov, Nikolay Kobyshev, Kenneth Vanhoey, Radu Timofte,and Luc Van Gool. DSLR-quality photos on mobile devices with deepconvolutional networks. arXiv preprint arXiv:1704.02470, 2017.

[231] Zongqing Lu, Noor Felemban, Kevin Chan, and Thomas La Porta.Demo abstract: On-demand information retrieval from videos usingdeep learning in wireless networks. In Internet-of-Things Designand Implementation (IoTDI), 2017 IEEE/ACM Second InternationalConference on, pages 279–280, 2017.

[232] Jemin Lee, Jinse Kwon, and Hyungshin Kim. Reducing distractionof smartwatch users with deep learning. In Proceedings of the 18thInternational Conference on Human-Computer Interaction with MobileDevices and Services Adjunct, pages 948–953. ACM, 2016.

[233] Toan H Vu, Le Dung, and Jia-Ching Wang. Transportation modedetection on mobile devices using recurrent nets. In Proceedings of the2016 ACM on Multimedia Conference, pages 392–396. ACM, 2016.

[234] Shih-Hau Fang, Yu-Xaing Fei, Zhezhuang Xu, and Yu Tsao. Learningtransportation modes from smartphone sensors based on deep neuralnetwork. IEEE Sensors Journal, 17(18):6111–6118, 2017.

[235] Daniele Ravì, Charence Wong, Benny Lo, and Guang-Zhong Yang. Adeep learning approach to on-node sensor data analytics for mobile orwearable devices. IEEE journal of biomedical and health informatics,21(1):56–64, 2017.

[236] Riccardo Miotto, Fei Wang, Shuang Wang, Xiaoqian Jiang, and Joel TDudley. Deep learning for healthcare: review, opportunities andchallenges. Briefings in Bioinformatics, page bbx044, 2017.

[237] Charissa Ann Ronao and Sung-Bae Cho. Human activity recognitionwith smartphone sensors using deep learning neural networks. ExpertSystems with Applications, 59:235–244, 2016.

[238] Jindong Wang, Yiqiang Chen, Shuji Hao, Xiaohui Peng, and Lisha Hu.Deep learning for sensor-based activity recognition: A survey. arXivpreprint arXiv:1707.03502, 2017.

[239] Xukan Ran, Haoliang Chen, Zhenming Liu, and Jiasi Chen. Deliveringdeep learning to mobile devices via offloading. In Proceedings of theWorkshop on Virtual Reality and Augmented Reality Network, pages42–47. ACM, 2017.

[240] Vishakha V Vyas, KH Walse, and RV Dharaskar. A survey on humanactivity recognition using smartphone. International Journal, 5(3),2017.

[241] Heiga Zen and Andrew Senior. Deep mixture density networks foracoustic modeling in statistical parametric speech synthesis. In Acous-tics, Speech and Signal Processing (ICASSP), 2014 IEEE InternationalConference on, pages 3844–3848, 2014.

[242] Siri Team. Deep Learning for Siri’s Voice: On-device Deep Mix-ture Density Networks for Hybrid Unit Selection Synthesis. https://machinelearning.apple.com/2017/08/06/siri-voices.html, 2017. [On-line; accessed 16-Sep-2017].

[243] Jon Barker, Ricard Marxer, Emmanuel Vincent, and Shinji Watan-abe. The third ’CHiME’speech separation and recognition challenge:Dataset, task and baselines. In Automatic Speech Recognition andUnderstanding (ASRU), 2015 IEEE Workshop on, pages 504–511,2015.

[244] Kai Zhao, Sasu Tarkoma, Siyuan Liu, and Huy Vo. Urban humanmobility data mining: An overview. In Big Data (Big Data), 2016IEEE International Conference on, pages 1911–1920, 2016.

[245] Shixiong Xia, Yi Liu, Guan Yuan, Mingjun Zhu, and Zhaohui Wang.Indoor fingerprint positioning based on Wi-Fi: An overview. ISPRSInternational Journal of Geo-Information, 6(5):135, 2017.

[246] Pavel Davidson and Robert Piche. A survey of selected indoorpositioning methods for smartphones. IEEE Communications Surveys& Tutorials, 2016.


[247] Xi Ouyang, Chaoyun Zhang, Pan Zhou, and Hao Jiang. DeepSpace:An online deep learning framework for mobile big data to understandhuman mobility patterns. arXiv preprint arXiv:1610.07009, 2016.

[248] Cheng Yang, Maosong Sun, Wayne Xin Zhao, Zhiyuan Liu, andEdward Y Chang. A neural network approach to jointly modeling socialnetworks and mobile trajectories. ACM Transactions on InformationSystems (TOIS), 35(4):36, 2017.

[249] Xuan Song, Hiroshi Kanasugi, and Ryosuke Shibasaki. DeepTransport:Prediction and simulation of human mobility and transportation modeat a citywide level. In IJCAI, pages 2618–2624, 2016.

[250] Junbo Zhang, Yu Zheng, and Dekang Qi. Deep spatio-temporalresidual networks for citywide crowd flows prediction. In Thirty-First Association for the Advancement of Artificial Intelligence (AAAI)Conference on Artificial Intelligence, 2017.

[251] J Venkata Subramanian and M Abdul Karim Sadiq. Implementationof artificial neural network for mobile movement prediction. IndianJournal of science and Technology, 7(6):858–863, 2014.

[252] Longinus S Ezema and Cosmas I Ani. Artificial neural networkapproach to mobile location estimation in gsm network. InternationalJournal of Electronics and Telecommunications, 63(1):39–44, 2017.

[253] Xuyu Wang, Lingjun Gao, Shiwen Mao, and Santosh Pandey. DeepFi:Deep learning for indoor fingerprinting using channel state information.In Wireless Communications and Networking Conference (WCNC),2015 IEEE, pages 1666–1671, 2015.

[254] Xuyu Wang, Xiangyu Wang, and Shiwen Mao. CiFi: Deep convolu-tional neural networks for indoor localization with 5 GHz Wi-Fi. In2017 IEEE International Conference on Communications (ICC), pages1–6.

[255] Xuyu Wang, Lingjun Gao, and Shiwen Mao. BiLoc: Bi-modal deeplearning for indoor localization with commodity 5GHz WiFi. IEEEAccess, 5:4209–4220, 2017.

[256] Michał Nowicki and Jan Wietrzykowski. Low-effort place recognitionwith WiFi fingerprints using deep learning. In International ConferenceAutomation, pages 575–584. Springer, 2017.

[257] Xiao Zhang, Jie Wang, Qinghua Gao, Xiaorui Ma, and Hongyu Wang.Device-free wireless localization and activity recognition with deeplearning. In Pervasive Computing and Communication Workshops(PerCom Workshops), 2016 IEEE International Conference on, pages1–5, 2016.

[258] Jie Wang, Xiao Zhang, Qinhua Gao, Hao Yue, and Hongyu Wang.Device-free wireless localization and activity recognition: A deeplearning approach. IEEE Transactions on Vehicular Technology,66(7):6258–6267, 2017.

[259] Mehdi Mohammadi, Ala Al-Fuqaha, Mohsen Guizani, and Jun-SeokOh. Semi-supervised deep reinforcement learning in support of IoTand smart city services. IEEE Internet of Things Journal, 2017.

[260] Anil Kumar Tirumala Ravi Kumar, Bernd Schäufele, Daniel Becker,Oliver Sawade, and Ilja Radusch. Indoor localization of vehicles usingdeep learning. In 2016 IEEE 17th International Symposium on Worldof Wireless, Mobile and Multimedia Networks (WoWMoM), pages 1–6.

[261] Zejia Zhengj and Juyang Weng. Mobile device based outdoor nav-igation with on-line learning neural network: A comparison withconvolutional neural network. In Proceedings of the IEEE Conferenceon Computer Vision and Pattern Recognition Workshops, pages 11–18,2016.

[262] Wei Zhang, Kan Liu, Weidong Zhang, Youmei Zhang, and Jason Gu.Deep neural networks for wireless localization in indoor and outdoorenvironments. Neurocomputing, 194:279–287, 2016.

[263] Joao Vieira, Erik Leitinger, Muris Sarajlic, Xuhong Li, and FredrikTufvesson. Deep convolutional neural networks for massive MIMOfingerprint-based positioning. In 28th Annual IEEE InternationalSymposium on Personal, Indoor and Mobile Radio Communications(IEEE PIMRC 2017). IEEE–Institute of Electrical and ElectronicsEngineers Inc., 2017.

[264] Jiang Xiao, Kaishun Wu, Youwen Yi, and Lionel M Ni. FIFS:Fine-grained indoor fingerprinting system. In 2012 21st InternationalConference on Computer Communications and Networks (ICCCN),pages 1–7.

[265] Moustafa Youssef and Ashok Agrawala. The Horus WLAN locationdetermination system. In Proceedings of the 3rd international confer-ence on Mobile systems, applications, and services, pages 205–218.ACM, 2005.

[266] Mauro Brunato and Roberto Battiti. Statistical learning theory for loca-tion fingerprinting in wireless LANs. Computer Networks, 47(6):825–845, 2005.

[267] Po-Jen Chuang and Yi-Jun Jiang. Effective neural network-based nodelocalisation scheme for wireless sensor networks. IET Wireless SensorSystems, 4(2):97–103, 2014.

[268] Marcin Bernas and Bartłomiej Płaczek. Fully connected neural net-works ensemble with signal strength clustering for indoor localizationin wireless sensor networks. International Journal of DistributedSensor Networks, 11(12):403242, 2015.

[269] Ashish Payal, Chandra Shekhar Rai, and BV Ramana Reddy. Analysisof some feedforward artificial neural network training algorithmsfor developing localization framework in wireless sensor networks.Wireless Personal Communications, 82(4):2519–2536, 2015.

[270] Yuhan Dong, Zheng Li, Rui Wang, and Kai Zhang. Range-basedlocalization in underwater wireless sensor networks using deep neuralnetwork. In IPSN, pages 321–322, 2017.

[271] Xiaofei Yan, Hong Cheng, Yandong Zhao, Wenhua Yu, Huan Huang,and Xiaoliang Zheng. Real-time identification of smoldering andflaming combustion phases in forest using a wireless sensor network-based multi-sensor system and artificial neural network. Sensors,16(8):1228, 2016.

[272] Baowei Wang, Xiaodu Gu, Li Ma, and Shuangshuang Yan. Temper-ature error correction based on BP neural network in meteorologicalwireless sensor network. International Journal of Sensor Networks,23(4):265–278, 2017.

[273] Ki-Seong Lee, Sun-Ro Lee, Youngmin Kim, and Chan-Gun Lee.Deep learning–based real-time query processing for wireless sensornetwork. International Journal of Distributed Sensor Networks,13(5):1550147717707896, 2017.

[274] Jiakai Li and Gursel Serpen. Adaptive and intelligent wireless sensornetworks through neural networks: an illustration for infrastructureadaptation through hopfield network. Applied Intelligence, 45(2):343–362, 2016.

[275] Fereshteh Khorasani and Hamid Reza Naji. Energy efficient data aggre-gation in wireless sensor networks using neural networks. InternationalJournal of Sensor Networks, 24(1):26–42, 2017.

[276] Chunlin Li, Xiaofu Xie, Yuejiang Huang, Hong Wang, and ChangxiNiu. Distributed data mining based on deep neural network for wirelesssensor network. International Journal of Distributed Sensor Networks,11(7):157453, 2015.

[277] Jonathan Ho and Stefano Ermon. Generative adversarial imitationlearning. In Advances in Neural Information Processing Systems, pages4565–4573, 2016.

[278] Michele Zorzi, Andrea Zanella, Alberto Testolin, Michele De FilippoDe Grazia, and Marco Zorzi. COBANETS: A new paradigm forcognitive communications systems. In Computing, Networking andCommunications (ICNC), 2016 International Conference on, pages 1–7. IEEE, 2016.

[279] Mehdi Roopaei, Paul Rad, and Mo Jamshidi. Deep learning controlfor complex and large scale cloud systems. Intelligent Automation &Soft Computing, pages 1–3, 2017.

[280] Shivashankar Subramanian and Arindam Banerjee. Poster: Deeplearning enabled M2M gateway for network optimization. In Proceed-ings of the 14th Annual International Conference on Mobile Systems,Applications, and Services Companion, pages 144–144. ACM, 2016.

[281] Ying He, Chengchao Liang, F Richard Yu, Nan Zhao, and Hongxi Yin.Optimization of cache-enabled opportunistic interference alignmentwireless networks: A big data deep reinforcement learning approach. In2017 IEEE International Conference on Communications (ICC), pages1–6.

[282] Ying He, Zheng Zhang, F Richard Yu, Nan Zhao, Hongxi Yin, Vic-tor CM Leung, and Yanhua Zhang. Deep reinforcement learning-basedoptimization for cache-enabled opportunistic interference alignmentwireless networks. IEEE Transactions on Vehicular Technology, 2017.

[283] Faris B Mismar and Brian L Evans. Deep reinforcement learningfor improving downlink mmwave communication performance. arXivpreprint arXiv:1707.02329, 2017.

[284] YangMin Lee. Classification of node degree based on deep learning androuting method applied for virtual route assignment. Ad Hoc Networks,58:70–85, 2017.

[285] Hua Yang, Zhimei Li, and Zhiyong Liu. Neural networks for MANETAODV: an optimization approach. Cluster Computing, pages 1–9,2017.

[286] Fengxiao Tang, Bomin Mao, Zubair Md Fadlullah, Nei Kato, OsamuAkashi, Takeru Inoue, and Kimihiro Mizutani. On removing routingprotocol from future wireless networks: A real-time deep learningapproach for intelligent traffic control. IEEE Wireless Communications,2017.


[287] Qingchen Zhang, Man Lin, Laurence T Yang, Zhikui Chen, and PengLi. Energy-efficient scheduling for real-time systems based on deep Q-learning model. IEEE Transactions on Sustainable Computing, 2017.

[288] Ribal Atallah, Chadi Assi, and Maurice Khabbaz. Deep reinforcementlearning-based scheduling for roadside communication networks. InModeling and Optimization in Mobile, Ad Hoc, and Wireless Networks(WiOpt), 2017 15th International Symposium on, pages 1–8. IEEE,2017.

[289] Sandeep Chinchali, Pan Hu, Tianshu Chu, Manu Sharma, Manu Bansal,Rakesh Misra, Marco Pavone, and Katti Sachin. Cellular network trafficscheduling with deep reinforcement learning. In National Conferenceon Artificial Intelligence (AAAI), 2018.

[290] Haoran Sun, Xiangyi Chen, Qingjiang Shi, Mingyi Hong, Xiao Fu,and Nikos D Sidiropoulos. Learning to optimize: Training deepneural networks for wireless resource management. arXiv preprintarXiv:1705.09412, 2017.

[291] Zhiyuan Xu, Yanzhi Wang, Jian Tang, Jing Wang, and Mustafa CenkGursoy. A deep reinforcement learning based framework for power-efficient resource allocation in cloud RANs. In 2017 IEEE Interna-tional Conference on Communications (ICC), pages 1–6.

[292] Paulo Victor R Ferreira, Randy Paffenroth, Alexander M Wyglinski,Timothy M Hackett, Sven G Bilén, Richard C Reinhart, and Dale JMortensen. Multi-objective reinforcement learning-based deep neuralnetworks for cognitive space communications. In Cognitive Commu-nications for Aerospace Applications Workshop (CCAA), 2017, pages1–8. IEEE, 2017.

[293] Hao Ye and Geoffrey Ye Li. Deep reinforcement learning for resourceallocation in v2v communications. arXiv preprint arXiv:1711.00968,2017.

[294] Oshri Naparstek and Kobi Cohen. Deep multi-user reinforcement learn-ing for dynamic spectrum access in multichannel wireless networks.arXiv preprint arXiv:1704.02613, 2017.

[295] Timothy J O’Shea and T Charles Clancy. Deep reinforcement learningradio control and signal detection with KeRLym, a gym RL agent.arXiv preprint arXiv:1605.09221, 2016.

[296] Michael Andri Wijaya, Kazuhiko Fukawa, and Hiroshi Suzuki.Intercell-interference cancellation and neural network transmit poweroptimization for MIMO channels. In Vehicular Technology Conference(VTC Fall), 2015 IEEE 82nd, pages 1–5, 2015.

[297] Michael Andri Wijaya, Kazuhiko Fukawa, and Hiroshi Suzuki. Neuralnetwork based transmit power control and interference cancellation forMIMO small cell networks. IEICE Transactions on Communications,99(5):1157–1169, 2016.

[298] Hongzi Mao, Ravi Netravali, and Mohammad Alizadeh. Neural adap-tive video streaming with pensieve. In Proceedings of the Conferenceof the ACM Special Interest Group on Data Communication, pages197–210. ACM, 2017.

[299] Tetsuya Oda, Ryoichiro Obukata, Makoto Ikeda, Leonard Barolli, andMakoto Takizawa. Design and implementation of a simulation systembased on deep Q-network for mobile actor node control in wirelesssensor and actor networks. In Advanced Information Networking andApplications Workshops (WAINA), 2017 31st International Conferenceon, pages 195–200. IEEE, 2017.

[300] Tetsuya Oda, Donald Elmazi, Miralda Cuka, Elis Kulla, Makoto Ikeda,and Leonard Barolli. Performance evaluation of a deep Q-networkbased simulation system for actor node mobility control in wirelesssensor and actor networks considering three-dimensional environment.In International Conference on Intelligent Networking and Collabora-tive Systems, pages 41–52. Springer, 2017.

[301] Hye-Young Kim and Jong-Min Kim. A load balancing scheme basedon deep-learning in IoT. Cluster Computing, 20(1):873–878, 2017.

[302] Lu Liu, Yu Cheng, Lin Cai, Sheng Zhou, and Zhisheng Niu. Deeplearning based optimization in wireless network. In 2017 IEEEInternational Conference on Communications (ICC), pages 1–6.

[303] Tom Schaul, John Quan, Ioannis Antonoglou, and David Silver. Pri-oritized experience replay. arXiv preprint arXiv:1511.05952, 2015.

[304] Qingjiang Shi, Meisam Razaviyayn, Zhi-Quan Luo, and Chen He. Aniteratively weighted MMSE approach to distributed sum-utility maxi-mization for a MIMO interfering broadcast channel. IEEE Transactionson Signal Processing, 59(9):4331–4340, 2011.

[305] Anna L Buczak and Erhan Guven. A survey of data mining andmachine learning methods for cyber security intrusion detection. IEEECommunications Surveys & Tutorials, 18(2):1153–1176, 2016.

[306] Mahmood Yousefi-Azar, Vijay Varadharajan, Len Hamey, and UdayTupakula. Autoencoder-based feature learning for cyber securityapplications. In Neural Networks (IJCNN), 2017 International JointConference on, pages 3854–3861. IEEE, 2017.

[307] Muhamad Erza Aminanto and Kwangjo Kim. Detecting impersonationattack in WiFi networks using deep learning approach. In Interna-tional Workshop on Information Security Applications, pages 136–147.Springer, 2016.

[308] Qingsong Feng, Zheng Dou, Chunmei Li, and Guangzhen Si. Anomalydetection of spectrum in wireless communication via deep autoencoder.In International Conference on Computer Science and its Applications,pages 259–265. Springer, 2016.

[309] Muhammad Altaf Khan, Shafiullah Khan, Bilal Shams, and JaimeLloret. Distributed flood attack detection mechanism using artificialneural network in wireless mesh networks. Security and Communica-tion Networks, 9(15):2715–2729, 2016.

[310] Abebe Abeshu Diro and Naveen Chilamkurti. Distributed attackdetection scheme using deep learning approach for Internet of Things.Future Generation Computer Systems, 2017.

[311] Alan Saied, Richard E Overill, and Tomasz Radzik. Detection ofknown and unknown DDoS attacks using artificial neural networks.Neurocomputing, 172:385–393, 2016.

[312] Manuel Lopez-Martin, Belen Carro, Antonio Sanchez-Esguevillas, andJaime Lloret. Conditional variational autoencoder for prediction andfeature recovery applied to intrusion detection in IoT. Sensors,17(9):1967, 2017.

[313] Zhenlong Yuan, Yongqiang Lu, Zhaoguo Wang, and Yibo Xue. Droid-Sec: deep learning in Android malware detection. In ACM SIGCOMMComputer Communication Review, volume 44, pages 371–372. ACM,2014.

[314] Zhenlong Yuan, Yongqiang Lu, and Yibo Xue. Droiddetector: Androidmalware characterization and detection using deep learning. TsinghuaScience and Technology, 21(1):114–123, 2016.

[315] Xin Su, Dafang Zhang, Wenjia Li, and Kai Zhao. A deep learningapproach to Android malware feature learning and detection. InTrustcom/BigDataSE/ISPA, 2016 IEEE, pages 244–251, 2016.

[316] Shifu Hou, Aaron Saas, Lifei Chen, and Yanfang Ye. Deep4MalDroid:A deep learning framework for Android malware detection based onlinux kernel system call graphs. In Web Intelligence Workshops (WIW),IEEE/WIC/ACM International Conference on, pages 104–111, 2016.

[317] Fabio Martinelli, Fiammetta Marulli, and Francesco Mercaldo. Eval-uating convolutional neural network for effective mobile malwaredetection. Procedia Computer Science, 112:2372–2381, 2017.

[318] Niall McLaughlin, Jesus Martinez del Rincon, BooJoong Kang,Suleiman Yerima, Paul Miller, Sakir Sezer, Yeganeh Safaei, ErikTrickel, Ziming Zhao, Adam Doupé, et al. Deep android malwaredetection. In Proceedings of the Seventh ACM on Conference on Dataand Application Security and Privacy, pages 301–308. ACM, 2017.

[319] Yuanfang Chen, Yan Zhang, and Sabita Maharjan. Deep learning forsecure mobile edge computing. arXiv preprint arXiv:1709.08025, 2017.

[320] Milan Oulehla, Zuzana Komínková Oplatková, and David Malanik.Detection of mobile botnets using neural networks. In Future Tech-nologies Conference (FTC), pages 1324–1326. IEEE, 2016.

[321] Pablo Torres, Carlos Catania, Sebastian Garcia, and Carlos GarciaGarino. An analysis of recurrent neural networks for botnet detectionbehavior. In Biennial Congress of Argentina (ARGENCON), 2016IEEE, pages 1–6, 2016.

[322] Meisam Eslahi, Moslem Yousefi, Maryam Var Naseri, YM Yussof,NM Tahir, and H Hashim. Mobile botnet detection model based onretrospective pattern recognition. International Journal of Security andIts Applications, 10(9):39–+, 2016.

[323] Mohammad Alauthaman, Nauman Aslam, Li Zhang, Rafe Alasem,and MA Hossain. A p2p botnet detection scheme based on decisiontree and adaptive multilayer neural networks. Neural Computing andApplications, pages 1–14, 2016.

[324] Reza Shokri and Vitaly Shmatikov. Privacy-preserving deep learning.In Proceedings of the 22nd ACM SIGSAC conference on computer andcommunications security, pages 1310–1321. ACM, 2015.

[325] Yoshinori Aono, Takuya Hayashi, Lihua Wang, Shiho Moriai, et al.Privacy-preserving deep learning: Revisited and enhanced. In Inter-national Conference on Applications and Techniques in InformationSecurity, pages 100–110. Springer, 2017.

[326] Seyed Ali Ossia, Ali Shahin Shamsabadi, Ali Taheri, Hamid R Rabiee,Nic Lane, and Hamed Haddadi. A hybrid deep learning architecture forprivacy-preserving mobile analytics. arXiv preprint arXiv:1703.02952,2017.

[327] Martín Abadi, Andy Chu, Ian Goodfellow, H Brendan McMahan, IlyaMironov, Kunal Talwar, and Li Zhang. Deep learning with differentialprivacy. In Proceedings of the 2016 ACM SIGSAC Conference onComputer and Communications Security, pages 308–318. ACM, 2016.


[328] Seyed Ali Osia, Ali Shahin Shamsabadi, Ali Taheri, Kleomenis Kat-evas, Hamed Haddadi, and Hamid R Rabiee. Private and scalablepersonal data analytics using a hybrid edge-cloud deep learning. IEEEComputer Magazine Special Issue on Mobile and Embedded DeepLearning, 2018.

[329] David M Blei, Andrew Y Ng, and Michael I Jordan. Latent dirichletallocation. Journal of machine Learning research, 3(Jan):993–1022,2003.

[330] Sandra Servia-Rodriguez, Liang Wang, Jianxin R Zhao, RichardMortier, and Hamed Haddadi. Personal model training under privacyconstraints. In Proceedings of the 3rd ACM/IEEE International Con-ference on Internet-of-Things Design and Implementation, Apr 2018.

[331] Briland Hitaj, Giuseppe Ateniese, and Fernando Perez-Cruz. Deepmodels under the GAN: Information leakage from collaborative deeplearning. arXiv preprint arXiv:1702.07464, 2017.

[332] Sam Greydanus. Learning the enigma with recurrent neural networks.arXiv preprint arXiv:1708.07576, 2017.

[333] Houssem Maghrebi, Thibault Portigliatti, and Emmanuel Prouff. Break-ing cryptographic implementations using deep learning techniques. InInternational Conference on Security, Privacy, and Applied Cryptog-raphy Engineering, pages 3–26. Springer, 2016.

[334] Donghwoon Kwon, Hyunjoo Kim, Jinoh Kim, Sang C Suh, IkkyunKim, and Kuinam J Kim. A survey of deep learning-based networkanomaly detection. Cluster Computing, pages 1–13, 2017.

[335] Mahbod Tavallaee, Ebrahim Bagheri, Wei Lu, and Ali A Ghorbani.A detailed analysis of the kdd cup 99 data set. In ComputationalIntelligence for Security and Defense Applications, 2009. CISDA 2009.IEEE Symposium on, pages 1–6, 2009.

[336] Kimberly Tam, Ali Feizollah, Nor Badrul Anuar, Rosli Salleh, andLorenzo Cavallaro. The evolution of android malware and Androidanalysis techniques. ACM Computing Surveys (CSUR), 49(4):76, 2017.

[337] Rafael A Rodríguez-Gómez, Gabriel Maciá-Fernández, and PedroGarcía-Teodoro. Survey and taxonomy of botnet research through life-cycle. ACM Computing Surveys (CSUR), 45(4):45, 2013.

[338] Menghan Liu, Haotian Jiang, Jia Chen, Alaa Badokhon, Xuetao Wei,and Ming-Chun Huang. A collaborative privacy-preserving deeplearning system in distributed mobile environment. In ComputationalScience and Computational Intelligence (CSCI), 2016 InternationalConference on, pages 192–197. IEEE, 2016.

[339] Sumit Chopra, Raia Hadsell, and Yann LeCun. Learning a similaritymetric discriminatively, with application to face verification. InComputer Vision and Pattern Recognition, 2005. CVPR 2005. IEEEComputer Society Conference on, volume 1, pages 539–546, 2005.

[340] Briland Hitaj, Paolo Gasti, Giuseppe Ateniese, and Fernando Perez-Cruz. PassGAN: A deep learning approach for password guessing.arXiv preprint arXiv:1709.00440, 2017.

[341] Neev Samuel, Tzvi Diskin, and Ami Wiesel. Deep MIMO detection.arXiv preprint arXiv:1706.01151, 2017.

[342] Xin Yan, Fei Long, Jingshuai Wang, Na Fu, Weihua Ou, and BinLiu. Signal detection of MIMO-OFDM system based on auto encoderand extreme learning machine. In Neural Networks (IJCNN), 2017International Joint Conference on, pages 1602–1606. IEEE, 2017.

[343] Timothy J O’Shea, Seth Hitefield, and Johnathan Corgan. End-to-endradio traffic sequence recognition with recurrent neural networks. InSignal and Information Processing (GlobalSIP), 2016 IEEE GlobalConference on, pages 277–281, 2016.

[344] Timothy J O’Shea, Kiran Karra, and T Charles Clancy. Learning tocommunicate: Channel auto-encoders, domain specific regularizers, andattention. In Signal Processing and Information Technology (ISSPIT),2016 IEEE International Symposium on, pages 223–228, 2016.

[345] Nathan E West and Tim O’Shea. Deep architectures for modulationrecognition. In Dynamic Spectrum Access Networks (DySPAN), 2017IEEE International Symposium on, pages 1–6, 2017.

[346] Sreeraj Rajendran, Wannes Meert, Domenico Giustiniano, VincentLenders, and Sofie Pollin. Distributed deep learning models for wirelesssignal classification with low-cost spectrum sensors. arXiv preprintarXiv:1707.08908, 2017.

[347] David Neumann, Wolfgang Utschick, and Thomas Wiese. Deepchannel estimation. In WSA 2017; 21th International ITG Workshopon Smart Antennas; Proceedings of, pages 1–6. VDE, 2017.

[348] Timothy J O’Shea, Tugba Erpek, and T Charles Clancy. Deep learningbased MIMO communications. arXiv preprint arXiv:1707.07980, 2017.

[349] Timothy J O’Shea, Latha Pemula, Dhruv Batra, and T Charles Clancy.Radio transformer networks: Attention models for learning to synchro-nize in wireless systems. In Signals, Systems and Computers, 201650th Asilomar Conference on, pages 662–666, 2016.

[350] Timothy J., O’Shea, and Jakob Hoydis. An introduction to deeplearning for the physical layer. arXiv preprint arXiv:1702.00832, 2017.

[351] Roberto Gonzalez, Alberto Garcia-Duran, Filipe Manco, MathiasNiepert, and Pelayo Vallina. Network data monetization using Net2Vec.In Proceedings of the SIGCOMM Posters and Demos, pages 37–39.ACM, 2017.

[352] Nichoas Kaminski, Irene Macaluso, Emanuele Di Pascale, AvishekNag, John Brady, Mark Kelly, Keith Nolan, Wael Guibene, and LindaDoyle. A neural-network-based realization of in-network computationfor the Internet of Things. In 2017 IEEE International Conference onCommunications (ICC), pages 1–6.

[353] Liang Xiao, Yanda Li, Guoan Han, Huaiyu Dai, and H Vincent Poor.A secure mobile crowdsensing game with deep reinforcement learning.IEEE Transactions on Information Forensics and Security, 2017.

[354] Sebastian Dörner, Sebastian Cammerer, Jakob Hoydis, and Stephan tenBrink. Deep learning-based communication over the air. arXiv preprintarXiv:1707.03384, 2017.

[355] Roberto Gonzalez, Filipe Manco, Alberto Garcia-Duran, Jose Mendes,Felipe Huici, Saverio Niccolini, and Mathias Niepert. Net2Vec: Deeplearning for the network. In Proceedings of the Workshop on Big DataAnalytics and Machine Learning for Data Communication Networks,pages 13–18, 2017.

[356] David H Wolpert and William G Macready. No free lunch theorems foroptimization. IEEE transactions on evolutionary computation, 1(1):67–82, 1997.

[357] Yu Cheng, Duo Wang, Pan Zhou, and Tao Zhang. A survey ofmodel compression and acceleration for deep neural networks. arXivpreprint arXiv:1710.09282, 2017. To appear in IEEE Signal ProcessingMagazine.

[358] Nicholas D Lane, Sourav Bhattacharya, Akhil Mathur, Petko Georgiev,Claudio Forlivesi, and Fahim Kawsar. Squeezing deep learning intomobile and embedded devices. IEEE Pervasive Computing, 16(3):82–88, 2017.

[359] Jie Tang, Dawei Sun, Shaoshan Liu, and Jean-Luc Gaudiot. Enablingdeep learning on IoT devices. Computer, 50(10):92–96, 2017.

[360] Forrest N Iandola, Song Han, Matthew W Moskewicz, Khalid Ashraf,William J Dally, and Kurt Keutzer. SqueezeNet: AlexNet-level accu-racy with 50x fewer parameters and< 0.5 MB model size. InternationalConference on Learning Representations (ICLR), 2017.

[361] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko,Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam.Mobilenets: Efficient convolutional neural networks for mobile visionapplications. arXiv preprint arXiv:1704.04861, 2017.

[362] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. ShuffleNet:An extremely efficient convolutional neural network for mobile devices.arXiv preprint arXiv:1707.01083, 2017.

[363] Qingchen Zhang, Laurence T Yang, Xingang Liu, Zhikui Chen, andPeng Li. A tucker deep computation model for mobile multimediafeature learning. ACM Transactions on Multimedia Computing, Com-munications, and Applications (TOMM), 13.

[364] Qingqing Cao, Niranjan Balasubramanian, and Aruna Balasubrama-nian. MobiRNN: Efficient recurrent neural network execution onmobile GPU. 2017.

[365] Chun-Fu Chen, Gwo Giun Lee, Vincent Sritapan, and Ching-YungLin. Deep convolutional neural network on iOS mobile devices. InSignal Processing Systems (SiPS), 2016 IEEE International Workshopon, pages 130–135, 2016.

[366] S Rallapalli, H Qiu, A Bency, S Karthikeyan, R Govindan, B Man-junath, and R Urgaonkar. Are very deep neural networks feasible onmobile devices. IEEE Trans. Circ. Syst. Video Technol, 2016.

[367] Nicholas D Lane, Sourav Bhattacharya, Petko Georgiev, ClaudioForlivesi, Lei Jiao, Lorena Qendro, and Fahim Kawsar. DeepX: Asoftware accelerator for low-power deep learning inference on mobiledevices. In Information Processing in Sensor Networks (IPSN), 201615th ACM/IEEE International Conference on, pages 1–12, 2016.

[368] Loc N Huynh, Rajesh Krishna Balan, and Youngki Lee. DeepMon:Building mobile GPU deep learning models for continuous vision ap-plications. In Proceedings of the 15th Annual International Conferenceon Mobile Systems, Applications, and Services, pages 186–186. ACM,2017.

[369] Jiaxiang Wu, Cong Leng, Yuhang Wang, Qinghao Hu, and Jian Cheng.Quantized convolutional neural networks for mobile devices. InProceedings of the IEEE Conference on Computer Vision and PatternRecognition, pages 4820–4828, 2016.

[370] Sourav Bhattacharya and Nicholas D Lane. Sparsification and sepa-ration of deep learning layers for constrained resource inference on


wearables. In Proceedings of the 14th ACM Conference on EmbeddedNetwork Sensor Systems CD-ROM, pages 176–189. ACM, 2016.

[371] Minsik Cho and Daniel Brand. Mec: Memory-efficient convolution fordeep neural network. arXiv preprint arXiv:1706.06873, 2017.

[372] Jia Guo and Miodrag Potkonjak. Pruning filters and classes: Towardson-device customization of convolutional neural networks. In Proceed-ings of the 1st International Workshop on Deep Learning for MobileSystems and Applications, pages 13–17. ACM, 2017.

[373] Shiming Li, Duo Liu, Chaoneng Xiang, Jianfeng Liu, Yingjian Ling,Tianjun Liao, and Liang Liang. Fitcnn: A cloud-assisted lightweightconvolutional neural network framework for mobile devices. In Em-bedded and Real-Time Computing Systems and Applications (RTCSA),2017 IEEE 23rd International Conference on, pages 1–6, 2017.

[374] Heiga Zen, Yannis Agiomyrgiannakis, Niels Egberts, Fergus Hender-son, and Przemysław Szczepaniak. Fast, compact, and high qualitylstm-rnn based statistical parametric speech synthesizers for mobiledevices. arXiv preprint arXiv:1606.06061, 2016.

[375] Gabriel Falcao, Luís A Alexandre, J Marques, Xavier Frazão, andJoao Maria. On the evaluation of energy-efficient deep learning usingstacked autoencoders on mobile gpus. In Parallel, Distributed andNetwork-based Processing (PDP), 2017 25th Euromicro InternationalConference on, pages 270–273. IEEE, 2017.

[376] Surat Teerapittayanon, Bradley McDanel, and HT Kung. Distributeddeep neural networks over the cloud, the edge and end devices. In Dis-tributed Computing Systems (ICDCS), 2017 IEEE 37th InternationalConference on, pages 328–339, 2017.

[377] Elias De Coninck, Tim Verbelen, Bert Vankeirsbilck, Steven Bohez,Pieter Simoens, Piet Demeester, and Bart Dhoedt. Distributed neuralnetworks for Internet of Things: the big-little approach. In Internet ofThings. IoT Infrastructures: Second International Summit, IoT 360Âr2015, Rome, Italy, October 27-29, 2015, Revised Selected Papers, PartII, pages 484–492. Springer, 2016.

[378] Shayegan Omidshafiei, Jason Pazis, Christopher Amato, Jonathan PHow, and John Vian. Deep decentralized multi-task multi-agentreinforcement learning under partial observability.

[379] Suyog Gupta, Wei Zhang, and Fei Wang. Model accuracy and runtimetradeoff in distributed deep learning: A systematic study. In DataMining (ICDM), 2016 IEEE 16th International Conference on, pages171–180, 2016.

[380] Benjamin Recht, Christopher Re, Stephen Wright, and Feng Niu.Hogwild: A lock-free approach to parallelizing stochastic gradientdescent. In Advances in neural information processing systems, pages693–701, 2011.

[381] Priya Goyal, Piotr Dollár, Ross Girshick, Pieter Noordhuis, LukaszWesolowski, Aapo Kyrola, Andrew Tulloch, Yangqing Jia, and Kaim-ing He. Accurate, large Minibatch SGD: Training imagenet in 1 hour.arXiv preprint arXiv:1706.02677, 2017.

[382] ShuaiZheng Ruiliang Zhang and JamesT Kwok. Asynchronous dis-tributed semi-stochastic gradient optimization. In AAAI, 2016.

[383] Corentin Hardy, Erwan Le Merrer, and Bruno Sericola. Distributeddeep learning on edge-devices: feasibility via adaptive compression.arXiv preprint arXiv:1702.04683, 2017.

[384] Brendan McMahan, Eider Moore, Daniel Ramage, Seth Hampson, andBlaise Aguera y Arcas. Communication-efficient learning of deep net-works from decentralized data. In Proceedings of the 20th InternationalConference on Artificial Intelligence and Statistics, volume 54, pages1273–1282, Fort Lauderdale, FL, USA, 20–22 Apr 2017.

[385] B McMahan and Daniel Ramage. Federated learning: Collaborativemachine learning without centralized training data. Google ResearchBlog, 2017.

[386] Keith Bonawitz, Vladimir Ivanov, Ben Kreuter, Antonio Marcedone,H. Brendan McMahan, Sarvar Patel, Daniel Ramage, Aaron Segal, andKarn Seth. Practical secure aggregation for privacy preserving machinelearning. Cryptology ePrint Archive, Report 2017/281, 2017. https://eprint.iacr.org/2017/281.

[387] Angelo Furno, Marco Fiore, and Razvan Stanica. Joint spatial andtemporal classification of mobile traffic demands. In Proc. IEEEINFOCOM, 2017.

[388] Zhiyuan Chen and Bing Liu. Lifelong machine learning. SynthesisLectures on Artificial Intelligence and Machine Learning, 10(3):1–145,2016.

[389] Sang-Woo Lee, Chung-Yeon Lee, Dong-Hyun Kwak, Jiwon Kim,Jeonghee Kim, and Byoung-Tak Zhang. Dual-memory deep learningarchitectures for lifelong learning of everyday human behaviors. InIJCAI, pages 1669–1675, 2016.

[390] Alex Graves, Greg Wayne, Malcolm Reynolds, Tim Harley, Ivo Dani-helka, Agnieszka Grabska-Barwinska, Sergio Gómez Colmenarejo,

Edward Grefenstette, Tiago Ramalho, John Agapiou, et al. Hybridcomputing using a neural network with dynamic external memory.Nature, 538(7626):471–476, 2016.

[391] German I Parisi, Jun Tani, Cornelius Weber, and Stefan Wermter.Lifelong learning of human actions with deep neural network self-organization. Neural Networks, 2017.

[392] Chen Tessler, Shahar Givony, Tom Zahavy, Daniel J Mankowitz, andShie Mannor. A deep hierarchical approach to lifelong learning inMinecraft. In AAAI, pages 1553–1561, 2017.

[393] Daniel López-Sánchez, Angélica González Arrieta, and Juan M Cor-chado. Deep neural networks and transfer learning applied to multime-dia web mining. In Distributed Computing and Artificial Intelligence,14th International Conference, volume 620, page 124. Springer, 2018.

[394] Ejder Bastug, Mehdi Bennis, and Mérouane Debbah. A transferlearning approach for cache-enabled wireless networks. In Modelingand Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOpt),2015 13th International Symposium on, pages 161–166. IEEE, 2015.

[395] Li Fei-Fei, Rob Fergus, and Pietro Perona. One-shot learning ofobject categories. IEEE transactions on pattern analysis and machineintelligence, 28(4):594–611, 2006.

[396] Mark Palatucci, Dean Pomerleau, Geoffrey E Hinton, and Tom MMitchell. Zero-shot learning with semantic output codes. In Advancesin neural information processing systems, pages 1410–1418, 2009.

[397] Oriol Vinyals, Charles Blundell, Tim Lillicrap, Daan Wierstra, et al.Matching networks for one shot learning. In Advances in NeuralInformation Processing Systems, pages 3630–3638, 2016.

[398] Soravit Changpinyo, Wei-Lun Chao, Boqing Gong, and Fei Sha.Synthesized classifiers for zero-shot learning. In Proceedings of theIEEE Conference on Computer Vision and Pattern Recognition, pages5327–5336, 2016.

[399] Junhyuk Oh, Satinder Singh, Honglak Lee, and Pushmeet Kohli. Zero-shot task generalization with multi-task deep reinforcement learning.arXiv preprint arXiv:1706.05064, 2017.

[400] Huandong Wang, Fengli Xu, Yong Li, Pengyu Zhang, and Depeng Jin.Understanding mobile traffic patterns of large scale cellular towers inurban environment. In Proc. ACM IMC, pages 225–238, 2015.

[401] Cristina Marquez, Marco Gramaglia, Marco Fiore, Albert Banchs,Cezary Ziemlicki, and Zbigniew Smoreda. Not all Apps are createdequal: Analysis of spatiotemporal heterogeneity in nationwide mobileservice usage. In Proceedings of the 13th ACM Conference onEmerging Networking Experiments and Technologies. ACM, 2017.

[402] Gianni Barlacchi, Marco De Nadai, Roberto Larcher, Antonio Casella,Cristiana Chitic, Giovanni Torrisi, Fabrizio Antonelli, AlessandroVespignani, Alex Pentland, and Bruno Lepri. A multi-source dataset ofurban life in the city of Milan and the province of Trentino. Scientificdata, 2, 2015.

[403] Liang Liu, Wangyang Wei, Dong Zhao, and Huadong Ma. Urbanresolution: New metric for measuring the quality of urban sensing.IEEE Transactions on Mobile Computing, 14(12):2560–2575, 2015.

[404] Denis Tikunov and Toshikazu Nishimura. Traffic prediction for mobilenetwork using Holt-Winters exponential smoothing. In Proc. SoftCOM,Split–fDubrovnik, Croatia, September 2007.

[405] Hyun-Woo Kim, Jun-Hui Lee, Yong-Hoon Choi, Young-Uk Chung, andHyukjoon Lee. Dynamic bandwidth provisioning using ARIMA-basedtraffic forecasting for Mobile WiMAX. Computer Communications,34(1):99–106, 2011.

[406] Muhammad Usama, Junaid Qadir, Aunn Raza, Hunain Arif, Kok-Lim Alvin Yau, Yehia Elkhatib, Amir Hussain, and Ala Al-Fuqaha.Unsupervised machine learning for networking: Techniques, applica-tions and research challenges. arXiv preprint arXiv:1709.06599, 2017.

[407] Chong Zhou and Randy C Paffenroth. Anomaly detection withrobust deep autoencoders. In Proceedings of the 23rd ACM SIGKDDInternational Conference on Knowledge Discovery and Data Mining,pages 665–674. ACM, 2017.

[408] Martín Abadi and David G Andersen. Learning to protect com-munications with adversarial neural cryptography. arXiv preprintarXiv:1610.06918, 2016.

[409] Silver David, Julian Schrittwieser, Karen Simonyan, IoannisAntonoglou, Aja Huang, Arthur Guez, Thomas Hubert, LucasBaker, Matthew Lai, Adrian Bolton, Yutian Chen, Timothy Lillicrap,Fan Hui, Laurent Sifre, George van den Driessche, Graepel Thore,and Demis Hassabis. Mastering the game of Go without humanknowledge. Nature, 550(7676):354–359, 2017.

[410] Jingjin Wu, Yujing Zhang, Moshe Zukerman, and Edward Kai-NingYung. Energy-efficient base-stations sleep-mode techniques in greencellular networks: A survey. IEEE communications surveys & tutorials,17(2):803–826, 2015.

Documents

IEEE COMMUNICATIONS SURVEYS & TUTORIALS 1 Deep … · 2018-03-13 · IEEE COMMUNICATIONS SURVEYS & TUTORIALS 1 Deep Learning in Mobile and Wireless Networking: A Survey Chaoyun Zhang,