5th International Conference on Advanced Technologies SCN-80

For Signal and Image Processing, ATSIP' 2020

September 02-05, 2020, Sfax, Tunisia

Deep learning for internet of things in fog computing:

Survey and Open Issues

Jihene Tmamna Emna Ben Ayed Mounir Ben Ayed

REGIM-Lab.:REsearch Groups in REGIM-Lab.:REsearch Groups in REGIM-Lab.:REsearch Groups in
Intelligent Machines Intelligent Machines Intelligent Machines
University of Sfax University of Sfax University of Sfax
National Engineering School of Sfax National Engineering School of Sfax Faculty of Science of Sfax
Email: [email protected] Email: [email protected] Email:[email protected]

Abstract— In recent years, the internet of things is getting very more than 50 billion, all connecting and communicating real
popular where it arose in several areas such as education, and time data on the internet by 2020 [5]. These devices will
healthcare to enhance our live. This popularity has led to an continuously generate intensives quantities of data which
increase number of IoT devices and thus generates massive will be referred as big IoT data.
volume of data. However, this data requires efficient methods
of analysis to provide intelligent services. Recently, the deep Thereafter, in order to explore these data and to make
learning can meet the requirements of IoT data analysis by important decisions, it is necessary to ensure effectively
providing techniques for large scale data analysis and analyzing the big IoT data. For that, the deep learning (DL) is
meaningful feature extraction. The deep learning considered as one of the important approaches, it can perform
implementation is traditionally delivered to cloud computing more reliable inference tasks [6]. Since, the DL can enhance
due to its high compute resources provisioning. However, given the performance of many applications when compared to
the sheer volume of IoT data, the cloud computing fall to meet traditional approaches [7] [8] [9]. However, due to its high
the IoT requirements, it presents many issues in term of time cost of computation and memory requirements the deep
response, large data transmission, energy consumption, etc. To
learning models are performed outside the edge devices.
address this challenges the fog computing, new layer between
cloud computing and internet of things devices, appears. So,
Meanwhile, to meet its computational requirements, the deep
moving the implementation of deep learning to fog computing learning models are implemented in the cloud which offers
can achieve the requirements of internet of things systems and tools to accelerate models for data analysis. However, the
enhance their performances. In this paper, we introduce deep transmission of the all data to the cloud can impact the
learning for internet of things, next the application of deep overall quality-of-experience (QoE) and quality of service
learning in internet of things. We address fog computing for the (QoS) for various IoT applications that are time-sensitive.
internet of things. Finally, we present the deep learning in fog
With respect to these problems, new emerging
computing.
technology appears that is fog computing. The fog computing
Keywords- internet of things, deep learning, fog computing dedicates the processing power of edge devices to provide
storage, computing, networking and processing services at
the edge of the internet [10] [11]. However, to ensure an IoT
I. INTRODUCTION big data analysis and fast response time, we see an emerging
trend toward embedded deploying of deep learning in fog
In recent decades, the world in changing through a new
computing. This will provide a real time analysis of the
technology that is internet of things (IoT). The IoT is a novel massive volume of data locally and improve the performance
paradigm that aims to combine the objects with internet and of IoT services.
data analysis tools to enhance our daily live [1] [2]. It
consists of interrelated devices like sensors which can The main scope of this paper is to introduce a
collect data, share information and communicate with each comprehensive survey in which we present the DL for IoT in
other using an underlying heterogeneous network such as the fog. Besides, we investigate the essence of fog computing
Bluetooth, WiFi and so forth [3]. Recently, The IoT has for IoT and the challenges of cloud computing to meet some
gained great attention from academic and industry in many IoT requirements.
aspects such as communication and networking [4]. In recent This paper is divided into five sections which are
years, internet of things has been observed in diversified organized as follows: Section II presents the DL for IoT and
areas including smart city, intelligent healthcare, smart provides an overview of the different applications related to
transportation, education and so on to provide services in DL in the IoT ecosystem. In section III, we introduce a
daily live. Due to its rapid adaptation, the IoT has witnessed survey of fog computing and we discuss some DL
a massive growth in the number of connected objects when applications in fog computing and the open issues. In section
it offers a universal accessibility of any given thing in IV, we introduce our proposition towards hybrid
anywhere at any time. At the edge of internet, the number of optimization approach for DL in fog computing. Finally,
IoT devices has growing exponentially, where Cisco section V closes the paper with conclusion.
predicted that the number of IoT connected devices will be

978-1-7281-7513-3/20/$31.00 ©2020 IEEE

Authorized licensed use limited to: Malaviya National Institute of Technology Jaipur. Downloaded on October 30,2023 at 18:13:50 UTC from IEEE Xplore. Restrictions apply.
 II. DEEP LEARNING FOR IOT abstraction [22]. So, DL is recently coming to play a
significant role in reaching solutions for IoT Data analytics.
Due to the increased demand for IoT application in
heterogeneous field, the number of connected things will C. Deep learning applications for IoT
increase and also will generate large quantities of data that The IoT and DL are two important technologies of this
introduce many new issues in data analysis following the era. As announced at Gartner Symposium/ITxpo 2016, these
characteristics of data. two technologies are among the top three strategic
A. Characteristics of the IoT data technology trends for 2017 [15] [23]. Deep learning
algorithms have been applied with successfully in several
The key characteristics of the IoT data are generally areas such as speech recognition, image classifications, video
expressed with the 5V (volume, velocity, value, variety, and recognition, audio processing and so on. There are several
veracity) [12]. Add to these characteristics, the IoT data has variants DL models such as Auto-encoders, Deep Belief
others features such as [13] [14]: Network, Recurrent Neural Networks, Long Short Term
• Large-Scale Data: a numberless of devices are Memory networks, Convolutional Neural Networks and
deployed for IoT applications. These devices will restricted Boltzmann Machines. Each model has better results
continuously generate a massive volume data which in special fields. For example, autoencoder is better in
should be analyzed and stored. As a result the dimensionality reduction for data visualization, data
volume of IoT data will continuously increase. denoising and anomaly detection and convolutional neural
networks work better in computer vision applications [17]
• Heterogeneity: due to the variety of IoT applications, and, while recurrent neural networks (RNN) provide superior
the IoT data sources are heterogeneous. Therefore, results in speech and natural language processing [24]. In this
similar to the data sources, IoT data are section, we present some IoT applications that take advantage
heterogeneous. of DL algorithms.
• Time and space correlation: each IoT object is Based on our observation, the most important IoT
located at a specific place and thus every IoT data applications is healthcare. It provides a remote monitoring of
has a location and time stamp. patient’s health condition through earlier disease discovery to
avoid the risks. Moreover, to control the human health state,
• Noise data: during the acquisition and transmission
there are various sensors to collect data and analysis tools to
of IoT data, quantities of noise and errors can occur
analyze the nature of healthcare problem [18]. Hence, DL
and can affect many of such data.
algorithm can handle complex health data and then extract
According to the aforementioned characteristics, the IoT hidden features. Several applications have been proposed
data complexity may lead to face many challenges such as using DL models for Disease Prediction. For example,
storage, processing and extraction of useful patterns or Alhussein et al. [25] developed a system to monitor epileptic
knowledge for future actions. Hence, the IoT data demand patients and to detect epileptic seizures. Their seizures
new and sophisticated data analysis approach to analyze the detection algorithm is based on convolutional neural
heterogeneous data in real time with high accuracy and networks and autoencoder to help identify patient’s state
efficiency. through their activities in which the input data is EEG signal.
Roopaei et al. [26] have also used the DL for face
B. IoT data and machine learning identification. They developed a facial perception platform to
Traditionally, to incorporate the intelligence into IoT help early stage Alzheimer’s patients which suffer from
application, many machine learning approaches were memory impaired.
proposed [15] [16]. However, based on the increasing The combination of DL with IoT has been also adopted in
number of IoT devices, the traditional machine learning smart city applications to provide a seamless and ubiquitous
algorithms are not efficient enough to deal with the new framework. Mohammadi et al. [27] propose model DL in
requirements of IoT systems [17]. Thus, it will be inadequate combination with other learning method reinforcement
to process IoT data. To address the new demands of IoT learning to predict the indoor localization. They use
system, DL approaches provide the solutions using many Variational Autoencoders (VAE) as the inference engine for
efficient learning models [18]. It is a branch of machine generalizing optimal. In addition, Kök et al. [28] used deep
learning techniques based on artificial neural networks, learning model and the IoT smart city data to reduce air
which extract high-level features [19]. pollution. They describe how predict future values of air
Recently, the DL has witnessed a great interest by quality in a smart city using Long Short Term Memory
researchers where the attempts to incorporate DL into IoT (LSTM) networks. Furthermore, DL can be used also in self-
systems have shown greater results in a multitude of driving cars in performing many tasks, such as detecting
applications [20]. The reasons behind choosing DL network obstacles, pedestrians, traffic signs, etc [17].
within IoT applications are the better performance with large As mentioned, the IoT applications create intensive
IoT data scale and the extraction of hidden features. quantities of complex data. Where these applications
Meanwhile, it offers the most appropriate tools that enable integrated with DL to ensure a fast and stream analysis. Yet,
the rapid analysis of complex data [21]. On the other hand, none of the mentioned works took into account the necessity
deep learning algorithms ensure the extraction of complex to consider the real time IoT data analysis requirement. From
data representations automatically at high levels of our point of view, to ensure a real time response, it would be

Authorized licensed use limited to: Malaviya National Institute of Technology Jaipur. Downloaded on October 30,2023 at 18:13:50 UTC from IEEE Xplore. Restrictions apply.
 better to analysis the IoT data in fog computing near the end It has three layers: the IoT layer, the fog layer and the
devices. cloud layer. Typically, the fog layer is composed of one or
III. THE FOG COMPUTING FOR IOT more fog domain [36], when the fog domain is formed by a
set of geographically distributed nodes called “fog nodes”
In the past years, cloud computing has attracted great such as routers, switches, cameras or any such devices have
interest. It has been applied in many sciences and engineering the capacity of storage and processing. As for, the IoT layer
fields due to its high cost-flexibility and efficiency achieved is consisted by IoT devices and end users devices. The fog
through consolidation. The cloud computing is a centralized nodes are linked to IoT and cloud computing. It can
server which storage, network management and computing communicate with IoT devices and end users devices by
functions [29]. The use of cloud computing had become a various network, like WiFi, Bluetooth, or 4G, to offers
popular trends, there are several commercial products such as storage, computation, and network services [37] and can also
Google, Amazon and so on offer a platform for on-demand
be communicated with cloud using the Internet in order to
access to several services such as network, processing power,
benefit of the high storage and computing resources of cloud
storage and thereby from anywhere and at any time [30].
[37] [38].
With the emergence of IoT, most IoT devices are
characterized by limited resources in terms of energy, In this sense, the combination of IoT with fog computing
computation and storage. As a result, it is impossible to can meet the demands of real time IoT application compared
undertake complex computational tasks into IoT devices [31]. to the use of cloud computing [35] [39] in term of low
For that, the cloud computing is considered as a greater latency and response time. Where, it enables real-time
solution for computation offloading due to its on-demand and services by analysis the data near the end devices.
scalable nature [32]. It is a viable solution for IoT data
analysis. A. Deep learning in fog computing
However, with the explosively increase of IoT data; it has In recent past, the DL models with billions of parameters
become difficult to send the large quantities of data from IoT and multiple layers need high computation and storage
devices to the centralized cloud. As a result, the integration of resources. Therefore, the most of existing DL applications are
cloud computing with IoT may lead to face many challenges implemented in cloud computing due to its efficient resources
in terms of data security and privacy, issues related to provisioning. Fig.2 shows the traditional cloud IoT data
transferring delay, disruption in communication network, and analysis, where all the IoT data generated by the sensors is
so on [33]. Thereby, the cloud computing cannot satisfy the transferred to the cloud when the DL takes place for
QoS requirements related to data analysis for IoT processing. However, cloud may not be ideal for IoT
applications especially for real time application. Hence, to especially for some real time applications in cases we need
overcome the aforementioned problems, there has been a low latency.
need for looking ‘beyond the clouds’ towards the fog
computing [34]. Recently, to satisfy the low latency requirements of IoT
application, the DL being gradually tending towards the
In 2012, Cisco introduced the fog computing to solve the network edges (fog computing) to improve the quality of
problems related to IoT applications in conventional Cloud decision making.
computing [35]. The fog computing is a novel architecture
located at edge of the network where it extends the traditional
architecture of cloud computing [36]. It forms a new layer
between distributed IoT sensors/devices and traditional cloud
computing [35]. Fig. 1 shows the fog computing layers.

Figure 2. Tradional cloud IoT data analytics

Hence, the new emerging concept of deep learning in fog

computing is considered as a better solution located near the
end user where the IoT data are generated. It has inherent
advantages in many fields such as self-driving cars,
healthcare, smart city and many other sensible time-latency
applications when the response time is extremely critical.
Several research works [40] [41] [42] [43] have focused on
Figure 1. Fog computing layer implementation of deep learning on fog computing to solve

Authorized licensed use limited to: Malaviya National Institute of Technology Jaipur. Downloaded on October 30,2023 at 18:13:50 UTC from IEEE Xplore. Restrictions apply.
 the cloud issues and to provide other benefits in terms of B. Discussion and open issues
bandwidth, real-time decisions and decreasing latency The main scope of this paper is to present an investigation
issues. While the fog analysis can provide the latency, there of DL in fog computing for IoT. From our point of view, the
are still several problems for adopt DL approaches in the fog combination between fog and deep learning approach needs
computing. One of the most challenges is that the fog nodes to be suitable for real-time IoT applications. However, Fog
are smaller and less powerful compared to cloud servers computing is at the initial phase, there some challenges.
[44], they do not have enough resources in terms of memory
size, storage, network, processing capability and The following units summarize the main issues that
computation. Thereby, the fog nodes are not sufficient deserve further investigation:
enough to support complex data analytics [45]. So, deploy
• Compression model: in this paper, we observed that
DL models in resources constrained devices still an open the DL models are computationally intensive making
challenge. them difficult to deploy on fog node with limited
To solve these issues, various methods were presented in resources. Therefore, it would be interesting to
compress the model to facilitate the use of DL in
the literatures that bring the DL approaches to fog devices
resources constrained devices. In our point of view
with limited resources. In this section, we present three we need to use the two techniques: pruning and
methods: compression, offloading and partitioning. quantification. More works needs to be done to
a) Model compression: develop a reliable algorithm to compress the model
without any accuracy loss.
This method consists to reduce the DL model size in order
to reduce their storage and computational cost. The • DL partitioning: we need to partition the DL task into
compression method can be classified to three categories: sub-task based on the fog nodes resources
availability to distribute the workload to multiple fog
pruning, quantification and knowledge distillation. Some
nodes. Much effort needs to be done to divide the DL
works such [46] and [47] use the pruning technique which
model between heterogeneous fog nodes to process
consists to explore the redundancy in the parameters of the in parallel and reduce the communication cost.
model and remove the redundant and unimportant
parameter. Other works compress the original model using • DL distribution: distribute the DL sub-task between
the quantification technique. This technique consists to heterogeneous fog nodes rather than the cloud
reduce the number of bits required to represent each weight computing. Then a natural question is how to
in the model such as in [48] and [49]. In [50] they use the schedule the DL sub-task between the fog nodes?
knowledge distillation to compress the DL model, this • Results aggregation: How we can collect and
technique consists to shift knowledge from a large model aggregate the results of each fog node in order to take
witch is considered as a teacher model into a small model a decision?
witch is considered as a student model.
• Discrimination: the ability of the system to reduce
b) Offloading: the energy consumption and response time.
this method consists to offload the computational DL model
• Utility and usability: DL in fog computing can be
to powerful servers like edge servers. Han et al. [51] and
useful in IoT applications particularity time sensitive
Ran et al. [52], for example, offload the dL model to the
application like healthcare.
powerful helper (edge server) to analyze video data.
c) Distribution: IV. TOWARDS AN APPROACH FOR DEEP LEARNING IN FOG
COMPUTING
this method consists to partition the DL model into sub-task
and distribute across multiple edge devices such as in [53] To handle these issues, we intend to build a compounded
and [54]. approach based on a hybrid optimization solution capable of
In the following table (Table I), we summarize the presented performing DL in fog computing. We will have as objective
function the reduction of the resource consumption, latency
previous works.
and speeding up execution taking in account the required
optimization criteria: accuracy, memory, energy, latency and
TABLE I. SUMMARY OF THE PREVIOUS WORKS DEALING WITH DEEP
LEARNING ON FOG COMPUTING network bandwidth. The general workflow of our approach is
presented in Fig.3. As illustrated in the figure, the principle of
Ref Method Results Critics this solution is to compress the deep learning models using
[46], [47], Reduce Computational
[48], [49], Compression energy, requirement still high,
the two compression techniques and to distribute them
[50] memory loss accuracy between the nodes. This approach consists of three main
components.
Latency,
[51], [52] offloading Latency still high
accuracy The first component is the compression model formed by
the two techniques of compression pruning and quantification
Latency, high communication
[53], [54] partitioning
energy cost
which work together to reduce the storage and calculation
needs of the model deep learning without affecting their
precisions. First we prune the model to eliminate the least

Authorized licensed use limited to: Malaviya National Institute of Technology Jaipur. Downloaded on October 30,2023 at 18:13:50 UTC from IEEE Xplore. Restrictions apply.
 important parameters. The number of connections is also to REFERENCES
be optimized. After, we use quantization technique to reduce [1] F. Lopes, J. Ferreira, R. Jardim-Goncalves, C. Agostinho, “Semantic
the number of bits of each parameter (weight, bias) in the Maps for IoT Network Reorganization in face of Sensor
pruned model. Malfunctioning”, 2017 IEEE International Conference on Systems,
Man, and Cybernetics (SMC), October 2017.
The second component is the partitioning model which
[2] C. L. Rose Jaren, Eldridge Scott, “The internet of things: an overview
consists of partitioning the compressed model into different Understanding the issues and challenges of a more connected world,”
partitions. These partitions must be executed in parallel on Internet Soc., no. October, 2015.
neighboring nodes. [3] A. Al-Fuqaha, M. Guizani, M. Mohammadi, M. Aledhari, M. Ayyash,
The third component is deployment. After the “Internet of Things: A Survey on Enabling Technologies, Protocols,
and Applications”, IEEE communication surveys & tutorials, vol. 17,
partitioning, each partition will be containerized and no. 4, 2015
distributed across nodes using the kubernetes platform.
[4] I. Yaqoob, E. Ahmed, IAT. Hashem, AIA. Ahmed, A. Gani, M.
Imran, and M. Guizani, “Internet of Things Architecture: Recent
Advances, Taxonomy, Requirements, and Open Challenges”, IEEE
Wireless Communications, vol.24, no.3, pp: 10-16, 2017
[5] M. A. Burhanuddin, A. A-J. Mohammed, R. Ismail and H. Basiron,
“Internet of Things Architecture: Current Challenges and Future
Direction of Research,” International Journal of Applied Engineering
Research ISSN, vol. 12, pp. 11055-11061, 2017.
[6] M. Song, K. Zhong, J. Zhang, Y. Hu, D. Liu, W. Zhang, J. Wang and
T. Li, “In-situ AI: Towards Autonomous and Incremental Deep
Learning for IoT Systems”, 2018 IEEE International Symposium on
High Performance Computer Architecture, pp: 92-103, 2018
[7] D.Ravi, C.Wong, B.Lo, and G-Z.Yang, “A Deep Learning Approach
Figure 3. Hybrid optimization method to on-Node Sensor Data Analytics for Mobile or Wearable Devices”,
IEEE Journal of biomedical and health informatics, pp: 56-64, 2017
V. CONCLUSION [8] D. Ravi, C. Wong, B. Lo, and G.-Z. Yang, “Deep learning for human
activity recognition: A resource efficient implementation on low-
Based on the huge increase of IoT objects and the new power devices,” in Proc. 2016 IEEE 13th Int. Conf. Wearable
emerging requirements in term of real time analysis and so Implantable Body Sensor Netw., pp. 71–76, Jun. 2016.
on, the IoT need appropriate approaches. This article has [9] M. A. Alsheikh, A. Selim, D. Niyato, L. Doyle, S. Lin and H-P. Tan,
investigated the suitability of deep learning based on fog “Deep activity recognition models with triaxial accelerometers”, Nov
computing for emerging IoT. We have introduce first the 2015. [Online]. Available: http://arxiv.org/abs/1511.04664
advantages of deep learning in providing a solution to [10] R. K. Naha, S. Garg, D. Georgakopoulos, P. P. Jayaraman, L. Gao, Y.
improve the data analysis and resolve learning challenges Xiang, and R. Ranjan, “Fog Computing: Survey of Trends,
related to big IoT data. We have also presented same Architectures, Requirements, and Research Directions”, IEEE access,
applications of deep learning in IoT. With the delay- latency pp. 47980- 48009, 2018
sensitive IoT application, the QoS and QoE requirements [11] A. V. Dastjerdi, H. Gupta, R. N. Calheiros, S. K. Ghosh, and
cannot be address by the cloud due to the large latency and R. Buyya, “Fog computing: Principles, architectures, and
network congestion. By the new emerging technology “fog applications”, in Internet of Things: Principle & Paradigms. San
computing” that became a solution to handle the challenges Mateo, CA, USA: Morgan Kaufmann, 2016.
of cloud computing, we are able to reduce the transmission [12] L. Farhan, S. T. Shukur, A. E. Alissa, M. Alrweg, U. Raza, R. Kharel,
latency and energy consumption for IoT devices and enhance “A Survey on the Challenges and Opportunities of the Internet of
Things (IoT)”, Eleventh International Conference on Sensing
the time response. We presented also the combination of DL Technology (ICST), 2017
with fog computing infrastructure and its benefits in term of
[13] L. Wang, C. A. Alexander, “Big Data Analytics and Cloud Computing
big and fast data analysis. This integration as it has in Internet of Things”, American Journal of Information Science and
advantages; it has also many challenges related to Computer Engineering Vol. 2, No. 6, pp. 70-78, 2016.
compression and distribution of DL models between fog [14] M. Chen, S. Mao, Y. Liu, “Big data: a survey”, Mobile Networks and
nodes. Applications”, pp.171-209, 2014.
In future, we intend to address the mentioned challenges. [15] R.F. Molanes, K. Amarasinghe, J. J. Rodriguez-Andina, and M.
Additionally, we intend to enhance our novel proposition that Manic, “Deep Learning and Reconfigurable Platforms in the Internet
of Things”, IEEE Industrial Electronics Magazine, pp. 36-49, 2018
should be able to execute DL for internet of things in fog
computing. [16] C. Perera, S. Member, A. Zaslavsky, and P. Christen, “Context aware
computing for the Internet of Things: A survey,” Commun. Surveys
Tuts, vol. 16, no. 1, pp. 1–41, 2014.
ACKNOWLEDGMENT
[17] M. Mohammadi, A. Al-Fuqaha, S. Sorour, M. Guizani, “Deep
The authors would like to acknowledge the financial Learning for IoT Big Data and Streaming Analytics: A Survey”, IEEE
support of this research by grants from the ARUB program Communications Surveys & Tutorials, 2018
under the jurisdiction of the General Direction of Scientific [18] M. Irfan and I. A. Hameed, “Deep Learning based Classification for
Research (DGRST) (Tunisia). Healthcare Data Analysis System”, IEEE Network, 2017
[19] Z. Md. Fadlullah, F. Tang, B. Mao, N. Kato, O. Akashi, T. Inoue, K.
Mizutani, “State-of-the-Art Deep Learning: Evolving Machine
Intelligence Toward Tomorrow’s Intelligent Network Traffic Control

Authorized licensed use limited to: Malaviya National Institute of Technology Jaipur. Downloaded on October 30,2023 at 18:13:50 UTC from IEEE Xplore. Restrictions apply.
 Systems”, IEEE Communications Surveys & Tutorials, vol.19, no.4, [38] M. Aazam, E. N. Huh, “Fog computing: The cloud-iot/ioe middleware
pp. 2432-2455, 2017 paradigm”, IEEE Potentials, pp. 40-44, 2016.
[20] S. Yao, Y. Zhao, H. Shao, C. Zhang, A. Zhang, D. Liu, S. Liu, L. Su [39] S. Sarkar , S. Chatterjee, and S. Misra, “Assessment of the suitability
and T.F. Abdelzaher, “ApDeepSense: Deep Learning Uncertainty of fog computing in the context of internet of things”, IEEE
Estimation Without the Pain for IoT Applications”, IEEE 38th Transactions on Cloud Computing PP(99): 1–1, 2015
International Conference on Distributed Computing Systems, 2018
[40] Z. Md. Fadlullah, A. S. Khan Pathan and H. Gacanin, “On Delay-
[21] S. Yao, Y. Zhao, A. Zhang, S.Hu, H. Shao, C. Zhang, S. Lu, and Sensitive Healthcare Data Analytics at the Network Edge Based on
T.Abdelzaher, “Deep learning for the internet of things,” Computer, Deep Learning”, 14th international Wireless Communications &
vol. 51, 2018. Mobile Computing Conference (IWCMC), 2018
[22] M. M Najafabadi, F. Villanustre, T. M Khoshgoftaar, N. Seliya, R. [41] L. Li, K. Ota, and M. Dong, “Deep Learning for Smart Industry:
Wald and E. Muharemagic, “Deep learning applications and Efficient Manufacture Inspection System With Fog Computing”,
challenges in big data analytics”, journal of big data, 2015 IEEE Transactions on industrial informatics, vol. 14, no. 10, pp. 4665,
4673, 2018
[23] K. Panetta. (2016, Oct. 18).Gartner’s top 10 strategic technology
trends for 2017 .Smarter with [42] M. Ali, A. Anjum, M. U. Yaseen, A. R. Zamani, D. Balouek-Thomert,
Gartner.[Online].Available:https://www.gartner.com/smarterwithgartn O. Rana, and M. Parashar, “Edge Enhanced Deep Learning System for
er/gartnerstop-10-technology-trends 2017/ Large-scale Video Stream Analytics”, IEEE 2nd International
Conference on Fog and Edge computing (ICFEC), 2018
[24] T. Gokmen, O. M. Onen, W. Haensch, “Training Deep Convolutional
Neural Networks with Resistive Cross-Point Devices”, Frontiers in [43] S. Teerapittayanon, B. McDanel, H. T. Kung, “Distributed Deep
neuroscience, vol.11, 2017 Neural Networks over the Cloud, the Edge and End Devices”, IEEE
37th International Conference on Distributed Computing System
[25] M. Alhussein, G. Muhammad, M. S. Hossain, S.U. Amin, “Cognitive (ICDCS) , 2017
IoT-Cloud Integration for Smart Healthcare: “Case Study for Epileptic
Seizure Detection and Monitoring”, Mobile Networks and [44] Z. Hao, E. Novak, S. Yi and Q. Li, “Challenges and Software
Applications, 2018 Architecture for Fog Computing”, IEEE Internet Computing, vol. 21,
no 2, pp. 44-53, 2017
[26] M. Roopaei, P. Rad, and J. J Prevost, “A Wearable IoT with Complex
Artificial Perception Embedding for Alzheimer Patients”, 2018 World [45] J. He, J. Wei, K. Chen, Z. Tang, Y. Zhou, and Y. Zhang, “Multi-tier
automation congress, 2018 Fog Computing with Large-scale IoT Data Analytics for Smart
Cities”, IEEE Internet of Things Journal, vol.5, no.2, pp. 677-686,
[27] M. Mohammadi, A. Al-Fuqaha, M. Guizani and J-S. Oh, “Semi- 2018
supervised Deep Reinforcement Learning in Support of IoT and Smart
City Services”, IEEE Internet of Things Journal, 2017 [46] F. Manessi, A. rozza, S. Bianco, P. Napoletana, R. Schettini.
[28] . Kök, M. U. imek and S. Özdemir, “A deep learning model for air “Automated pruning for deep neural network compression”, 24th
quality prediction in smart cities”, IEEE International Conference on International Conference on Pattern Recognition (ICPR). IEEE, p.
Big Data, 2017 657-664, 2018.
[29] Y. Ai, M. Peng, K. Zhang, “Edge Computing Technologies for [47] J. Kim, C. Park, H. J. Jung, Y. Choe “Differentiable Pruning Method
Internet of Things: A Primer”, Digital Communications and Networks, for Neural Networks”, arXiv preprint arXiv, pp. 1904-10921, 2019.
vol.4, no.2, pp. 77-86, 2018 [48] X. Zhu, W. Zhou and H. Li, “Adaptive layerwise quantization for
[30] M. Aazam, S. Zeadally, K. A. Harras, “Offloading in Fog Computing deep neural network compression”, IEEE International Conference on
for IoT: Review, Enabling Technologies, and Research Multimedia and Expo (ICME). IEEE, p. 1-6, 2018.
Opportunities”, Future Generation Computer Systems, 2018
[49] N. Floropoulos and A. Tefas, “Complete vector quantization of
[31] W.Yu, F.Liang, X.He, W.G. Hatcher, Ch.Lu, J.Liny, and X.Yangy, “A feedforward neural networks”, Neurocomputing, vol. 367, p. 55-63,
Survey on the Edge Computing for the Internet of Things”, IEEE 2019
access, vol:6, pp. 6900-6919, 2018
[50] B. Reddy, Y. H.Kim, S. Yun, C. Seo and J. Jang, “Real-time driver
[32] G.Premsankar, M.D.Francesco, and T.Taleb, “Edge Computing for the drowsiness detection for embedded system using model compression
Internet of Things: A Case Study”, IEEE Internet of Things Journal, of deep neural networks”, Proceedings of the IEEE Conference on
pp. 1275 – 1284, 2018 Computer Vision and Pattern Recognition Workshops, p. 121-128,
[33] S. B.Nath, H.Gupta, S.Chakraborty, S. K Ghosh, “A Survey of Fog 2017.
Computing and Communication: Current Researches and Future [51] S. Han, H. Shen, M. Philipose, S. Agarwal, A. Wolman, and A.
Directions”, Networking and Internet Architecture, 2018 Krishnamurthy, “MCDNN: An approximation-based execution
[34] B.Varghese, N.Wang, S.Barbhuiya, P.Kilpatrick and framework for deep stream processing under resource constraints,” in
D.S.Nikolopoulos, “Challenges and Opportunities in Edge Proc. ACM Mobisys, pp. 123–136, 2016.
Computing”, 2016 IEEE International Conference on Smart Cloud [52] X. Ran, H. Chen, X. Zhu, Z. Liu, and J. Chen, “DeepDecision: A
(SmartCloud), 2016. mobile deep learning framework for edge video analytics”, Proc.
[35] R.Mahmud, R.Kotagiri and R.Buyya, “Fog Computing: A Taxonomy, IEEE INFOCOM, pp. 1421–1429, 2018.
Survey and Future Directions” Internet of everything, Spring, [53] F. M. C. d. Oliveira and E. Borin, “Partitioning Convolutional Neural
Singapore, pp. 103-130, 2018 Networks to Maximize the Inference Rate on Constrained IoT
[36] C. Mouradian, D. Naboulsi, S. Yangui, R. H. Glitho, M. J. Morrow, Devices”, Future Internet, vol. 11, no 10, p. 209, 2019.
and P. A. Polakos, “A Comprehensive Survey on Fog Computing: [54] J. Mao, X. Chen, K. W. Nixon, C. Krieger, and Y. Chen, “MoDNN:
State-of-the-art and Research Challenges”, IEEE communications Local distributed mobile computing system for deep neural network”,
surveys & tutorials, pp. 416- 464, 2018 Design, Automation & Test in Europe Conference & Exhibition
[37] P. Hu, S. Dhelim, H. Ning, T. Qiu, “Survey on Fog Computing: (DATE), pp. 1396–1401, Mar. 2017.
Architecture, Key Technologies, Applications and Open Issues”,
Journal of Network and Computer Applications, 2017

Authorized licensed use limited to: Malaviya National Institute of Technology Jaipur. Downloaded on October 30,2023 at 18:13:50 UTC from IEEE Xplore. Restrictions apply.